Shifts in Care Utilization Patterns During the COVID-19 Pandemic: A High-Dimensional Study of Presentations for Ophthalmic Conditions in the US

doi:10.21203/rs.3.rs-1950480/v1

Download PDF

Article

Shifts in Care Utilization Patterns During the COVID-19 Pandemic: A High-Dimensional Study of Presentations for Ophthalmic Conditions in the US

https://doi.org/10.21203/rs.3.rs-1950480/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 14 Dec, 2023

Read the published version in Communications Medicine →

Version 1

posted

You are reading this latest preprint version

Background: Healthcare restrictions during the COVID-19 pandemic, particularly in ophthalmology, resulted in a differential underutilization of care with potential long-term consequences. An analytic approach is needed to characterize pandemic health services usage across many conditions.

Methods: A common analytical framework identified pandemic care utilization patterns across 261 ophthalmic diagnoses. Using a United States eye care registry, predictions of utilization expected without the pandemic were established for each diagnosis via models trained on pre-pandemic data. Pandemic effects on utilization were estimated by calculating deviations between observed and expected patient volumes from January 2020 to December 2021. Two sub-periods of focus were defined: the hiatus (March-May 2020), when many clinical visits were curtailed, and post-hiatus (June 2020-December 2021). Characteristics of conditions with similar deviation patterns were ascertained from cluster analyses, data visualizations, and hypothesis testing.

Results: Analyzing records from 44.62 million patients and 2,455 practices, we observed lasting reductions in ophthalmic care utilization, including visits for leading causes of visual impairment (age-related macular degeneration, diabetic retinopathy, cataract, glaucoma). Mean deviations among all diagnoses were 67% below expectation during the nadir of the hiatus (April 2020), and 13% post-hiatus. Less severe conditions experienced greater utilization reductions, with heterogeneities across diagnosis categories and pandemic phases. Intense post-hiatus utilization reductions occurred among non-vision-threatening conditions or asymptomatic precursors of vision-threatening diseases. Many conditions with above-average post-hiatus utilization posed a risk for irreversible morbidity, such as emergent pediatric, retinal, or uveitic diseases.

Conclusions: We derived high-resolution insights on pandemic care utilization from high-dimensional data using an analytical framework that can be applied to study healthcare disruptions in other settings and inform efforts to pinpoint unmet clinical needs.

Health sciences/Health care/Health services

Health sciences/Diseases/Eye diseases

Health sciences/Health care/Public health

During the COVID-19 pandemic, sustained disruptions to essential health services were documented worldwide.^1,2 In the United States, historic declines in care utilization were widely observed during the pandemic’s acute phase in Spring 2020. Despite subsequent recoveries in patient volume and healthcare spending, many measures of overall healthcare utilization failed to return to,^3,4 or exceed,^2,5 pre-pandemic levels. Among all specialties, ophthalmology has experienced one of the most severe disruptions to care,^4,6,7 likely related to the specialty’s reliance on close-proximity examinations, an older patient population, and a large proportion of elective procedures.^8,9 Studying the patterns of ophthalmic care utilization during the pandemic may therefore not only reveal emerging areas of potentially unmet need in population vision health, but also contribute further insights on key characteristics associated with services that patients and health systems may regard as priorities. We therefore conducted a high-dimensional study to understand shifts in eye care utilization patterns for a comprehensive set of ophthalmic diagnoses during the first two years of the COVID-19 pandemic in the United States.

Redistributions of essential healthcare resources during the acute phase of the pandemic were required to minimize mortality, but the patterns of, and reasons explaining, sustained utilization reductions in the post-acute phase are not entirely clear.¹⁰ Previous studies by health economists have formally estimated the responsiveness, or elasticity, of demand of healthcare services to changes in cost or income.^11–13 For instance, emergency room visits tended to exhibit little change in demand in response to changes in price, whereas pharmaceuticals, mental health/substance abuse treatment, and specialist care had high elasticities of demand.¹¹ Similarly, we explored how utilization levels for a wide range of ocular diagnoses exhibited varying degrees of sensitivity to possible pandemic-related restrictions to the seeking or delivery of care (e.g., resource constraints, behavioral modifications). We specifically investigated possible factors driving the differential underutilization of ophthalmic care during the pandemic by examining whether characteristics of medical problems themselves—namely, disease severity—were associated with observed changes in care utilization relative to levels expected in the absence of the pandemic.

Therefore, unlike previous studies^4,14–16 that investigated how pandemic utilization trends differed by patient demographic characteristics or the setting and type of care, or ones that studied overall categories of diagnoses or a limited selection of medical problems,¹⁷ our investigation explored the variations in healthcare usage across a comprehensive spectrum of detailed diagnoses from a single specialty via a common analytical framework that related care utilization patterns to specific attributes of medical problems. Using longitudinal electronic health record (EHR) data from a large national disease registry inclusive of most ambulatory ophthalmic care in the United States, we extracted time series data for each ocular condition included in the study, generated predictive models from pre-pandemic data to establish counterfactual levels of healthcare utilization for each condition expected in the absence of the pandemic, estimated pandemic effects on utilization by calculating deviations of observed data from predictions, and investigated characteristics of conditions with similar deviations. In doing so, we illustrate an analytic approach that can be adapted to different settings to serve as initial tool for an exploratory, but expansive and granular, assessment of impacts to care utilization during prolonged disruptions to healthcare access. The high-resolution insights generated from this high-dimensional analysis may inform studies aiming to determine the clinical impacts of missed or delayed care in specific patient populations and disease cohorts, monitor the pandemic’s effect on healthcare access, and clarify distinctions between harmful and benign reductions to care. Furthermore, because this study is, to our knowledge, the largest and most comprehensive characterization of ophthalmic care patterns during COVID-19, its findings may also contribute to a developing understanding of the pandemic’s long-term impact on population vision health.

Patterns in care utilization were characterized using a common analytical framework, which consisted of multiple stages (Figure 1).

Inventory of Diagnoses

We constructed an expansive inventory of ocular conditions to study by grouping more than 3,300 ophthalmic ICD-10 diagnosis codes into 336 clinically meaningful diagnosis entities adapted from categorizations provided by the US Agency for Healthcare Research and Quality Clinical Classifications Software Refined (CCSR) database version 2020.2.¹⁸ We followed a common set of considerations to modify these categorizations of ICD-10 codes where needed for the analytic purposes of this study, and to further create more granular groupings of these ICD-10 codes into diagnosis entities (Text S4). For instance, all diagnosis entities were required to have a sufficient utilization level (i.e., above single-digit patient counts) for each month of the study period (January 2017 to December 2021). After establishing an initial set of 336 diagnosis entities, to ensure that study findings were based on reliable predictions of utilization levels expected in the absence of the pandemic, conditions with poor counterfactual model performance (considered as ≥12.5% root-mean-squared percentage error [RMSPE]) were also excluded, resulting in a final set of 261 diagnosis entities attributed to 13 mutually exclusive diagnosis categories (e.g., blindness and vision defects, cataract and other lens disorders, corneal and external disease) (Table S1) included for analysis in this study.

Data Extraction and Study Setting

Monthly numbers of patients documented with each diagnosis entity were queried from the American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight), the United States’ first comprehensive clinical database of eye diseases, encompassing over 440 million patient visits from 73 million unique patients and 3,000 practices as of April 1, 2022. The IRIS Registry consists of deidentified data from EHRs submitted on a daily or weekly basis by practices representing over 70% of US ophthalmologists, and contains records on ocular diagnoses, procedures, visits, examinations, medications; patient demographic variables; and the setting of care. Although its coverage is not as large as with private, outpatient practices, the IRIS Registry, which is compatible with multiple EHR software vendors,¹⁹ is also integrated with approximately 33% of Association of University Professors of Ophthalmology member institutions. The methods and operations of the IRIS Registry have been previously described.²⁰

We defined a global study period of January 1, 2017, to December 31, 2021, and delineated a pre-pandemic period (2017-2019) that was used to establish counterfactual baselines for patient volumes, and a pandemic period (2020-2021) for which deviations from expectation were calculated. Within the pandemic period, we defined a hiatus period (March-May 2020) to highlight the steepest reductions in utilization during the acute phase of the pandemic, and a post-hiatus period (June 2020-December 2021) to represent the post-Spring 2020 recoveries that were observed widely in our analysis and elsewhere.^3–5 Patients included in the monthly count for a diagnosis entity were documented with a physician encounter (a procedure, ocular examination, or visit record) on the same date as the diagnosis. Furthermore, to exclude from consideration changes in utilization trends that may be attributable to new openings, permanent closures, or other changes in the data reporting statuses of practices, we only analyzed EHRs from practices that reported data to the IRIS Registry throughout all months of the global study period.

Establishing Counterfactual Levels of Pandemic Care Utilization

A parametric approach was used to estimate monthly numbers of patients expected for each diagnosis entity in the absence of the pandemic. We fit predictive models to time series data from the pre-pandemic period and used the fitted models to predict counterfactual utilization levels during the pandemic. In particular, for each condition, we used a leave-out-one-year blocked cross validation algorithm to select a best-fitting model, among a prespecified set of candidate models, that had the best predictive ability, which was determined by identifying the candidate model with the lowest average mean squared error across all holdout years from the pre-pandemic period.^21,22 Candidate predictive generalized linear models initially assumed conditionally Poisson-distributed outcomes and included combinations of terms for seasonality (monthly fixed effects and/or harmonic terms with 3-, 6-, and 12-month periodicities) and a linear trend over years ( Text S1). Finally, to establish counterfactual levels of care utilization during the pandemic period, errors in the fit of the model selected for each diagnosis entity were assessed for overdispersion, after which 100,000 Monte Carlo-simulated predictions were generated from conditional Poisson, over-dispersed Poisson, or negative binomial distributions in accordance with the assessed mean-variance relationship of model predictions to residual values. 95% prediction intervals (PIs) for monthly counterfactual predictions were subsequently calculated (Text S2). For each diagnosis entity, the specification of the selected model, along with model performance metrics and descriptive summary statistics, are provided in Table S2.

Measures of Pandemic Effects on Care Utilization and Time to Recovery

Our primary outcome, δ, was, for each diagnosis entity, its deviation from expected care utilization, calculated as the relative difference between observed and expected numbers of patients for a given month:

For each deviation, 95% confidence intervals (CIs) and two-sided empirical p-values were obtained by referencing the counterfactual distribution of the monthly prediction. We also estimated average deviations across summary timeframes: for each quarter (Q3 2020 to Q4 2021), and across all months, of the post-hiatus period. Average deviations were computed by taking the mean of the joint distribution of deviations across each month in the summary period, and 95% CIs and two-sided empirical p-values were similarly obtained.

As a secondary outcome, we examined the time it took for utilization to recover to expected values. Recovery was defined as three or more consecutive months for which no statistically significant negative deviations from expectation ( < 0, adjusted p ≤ 0.05) were recorded. Among conditions that recovered, “sustained recovery” described those that did not experience further significant negative deviations, and “partial recovery” for the conditions that did.

Analysis of Care Utilization Patterns and their Association with Diagnosis Severity

To identify groups of diagnosis entities with common longitudinal patterns in monthly deviations, we applied a hierarchical clustering algorithm (using the Euclidean distance function and the complete linkage method²³) and generated cluster heatmaps.^24–26 To examine associations between condition severities and utilization patterns, we relied on both qualitative insights derived from visualizations of condition deviations throughout the pandemic, and hypothesis tests that used disease severity measures provided by prior studies. Based on a global survey of over 70 experts on the severities of common ocular emergencies, Bourges et al. aggregated severity rankings for each surveyed condition, ranging from 1 (least severe) to 5 (most severe).²⁷ We mapped 36 diagnosis entities from our study to the set in Bourges et al. (Table S3), and examined associations between severity and utilization at different time points using univariable linear regressions. Similarly, using Wilcoxon rank-sum tests, we investigated whether significant differences in deviations existed between vision-threatening (VT) and non-vision-threatening (NVT) subtypes of the leading causes of visual impairment in the US:²⁸ diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma (Table S4).

P-values ≤ 0.05 were considered statistically significant. To control for multiple testing among estimated monthly or quarterly deviations for each diagnosis entity, we calculated false discovery rate (FDR)-adjusted p-values for all monthly and quarterly deviations using the Benjamini-Hochberg method, with the FDR threshold set at 0.05.²⁹ Data extraction was performed using Amazon Redshift version 1.0.38698 (PostgreSQL 8.0.2), and all statistical analyses and visualizations³⁰ were produced using R version 4.1.0.^31,32 To reduce concerns regarding our choice of a maximum allowable RMSPE, or the criteria we applied to identify diagnosis records and patients eligible for inclusion, we conducted sensitivity analyses to demonstrate no major changes in primary outcomes across different model performance thresholds or sets of inclusion criteria (Text S3).

Counterfactual expectations across all 261 diagnosis entities meeting study criteria were established using EHRs from 44.62 million unique patients and 2,455 practices, with a mean (SD) out-of-sample RMSPE of 10.3% (7.3%). Sustained reductions in care utilization were observed across nearly all diagnoses studied, including conditions that represent the least severe forms of the leading causes of treatable low vision and blindness in the US²⁸ (Figure 2): early-stage dry AMD, non-proliferative DR without diabetic macular edema, age-related cataract, and glaucoma suspect. Across all conditions, the sharpest decreases in utilization occurred during the hiatus (from March to May 2020, with a nadir in April 2020); and despite a rebound, utilization mostly (for 94.3% of diagnoses) remained below pre-pandemic volumes in the post-hiatus period (Figure 3A). On average, deviations were below expectation by 67% (14%) in the nadir of the hiatus (δ_H = -0.67), and by 13% (9%) post-hiatus (δ_P_H = -0.13) (Figure 3B).

Care Utilization Patterns in Relation to Disease Severity Despite an overall reduction in care utilization, we observed that decreases tended to be smaller for more severe conditions, suggesting a continued prioritization of care for diagnoses perceived as more urgent. This inverse relationship between condition severity and the magnitude of underutilization can be first observed among the 13 diagnosis categories encompassing all 261 diagnosis entities studied (Figure 3B). Throughout the pandemic study period, utilization reductions were less pronounced for the more severe diagnostic categories of ocular globe injuries/intraocular foreign bodies (OGI/IOFB) (mean within-category deviations: δ_H = -0.49, δ_P_H = -0.08), uveitis and ocular inflammation (-0.51, -0.12), and retinal and vitreous conditions (-0.60, -0.11), whereas the less severe categories of refractive error (-0.89, -0.14), strabismus (-0.84, -0.15), and blindness and vision defects (-0.78, -0.17) consistently experienced greater decreases in utilization.

Associations between severity and utilization were further observed among diagnosis entities themselves, but this relationship was not consistently present throughout different pandemic phases. Differences in utilization levels based on disease severity were conspicuous among nearly all diagnosis entities during the hiatus but were less conspicuous in the post-hiatus period for some groupings of conditions. Among the common eye conditions of DR, AMD, glaucoma, and cataract, deviations from expected utilization were clearly separated by condition severity in April 2020, but not during the post-hiatus period (Figure 4A). For instance, more severe stages of AMD were associated with lesser utilization reductions during the hiatus (δ_H = -0.32, -0.53, -0.75, -0.81 for exudative AMD with active neovascularization; exudative AMD without active neovascularization; early-stage nonexudative AMD; drusen of macula, respectively); but there were minimal differences among post-hiatus deviations (δ_P_H = -0.16, -0.10, -0.15, -0.17). On the other hand, the relative rankings of deviations remained consistent over time for other sets of conditions like those related to neuro-ophthalmic diseases, as reflected by a strong positive correlation between hiatus and post-hiatus deviations (r = 0.73, p = 0.001) (Figure 4B). Care utilization for oculomotor (δ_H = -0.52, δ_P_H = -0.07) and abducens (-0.59, -0.07) nerve palsies, and optic neuritis (-0.63, -0.06) experienced relatively limited decreases throughout both pandemic sub-periods, whereas pupillary function anomalies (-0.84, -0.19) and irregular eye movements (-0.83, -0.16) consistently exhibited greater utilization reductions. The presence of a strong positive correlation between hiatus and post-hiatus deviations suggests that conditions with greater reductions in care utilization during the hiatus period continued to exhibit relatively lesser rebounds in utilization post-hiatus. This could indicate that utilization for some categories of conditions continued to be sensitive to pandemic-related constraints to healthcare provision more so than other sets of diagnoses in the post-hiatus phase, but further research is needed to form a robust interpretation. Discrepancies in utilization levels based on condition severity, and relationships between hiatus and post-hiatus deviations, can be similarly observed among conditions in other diagnosis categories (Figures S1-S6).

Furthermore, among the 36 ocular emergencies we examined with an associated severity ranking, median deviations from expectation progressively increased with condition severities during both hiatus and post-hiatus periods (Figure 5A). For every unit increase in severity ranking, deviations increased, on average, by 5.5% (p < 0.001) during the hiatus, and 1.9% (p = 0.04) post-hiatus. Among sets of AMD, DR, and glaucoma diagnoses, NVT conditions consistently exhibited more reductions than their VT counterparts during April 2020 (p = 0.06, 0.03, 0.01, respectively), but this relationship was non-existent or weaker post-hiatus (p = 0.56, 0.89, 0.09) (Figure 5B).

Identification of Clusters of Diagnosis Entities with Similar Longitudinal Deviation Patterns

Patterns of longitudinal deviations in care utilization across all diagnoses are summarized using a cluster heatmap of quarterly post-hiatus deviations (Figure 6), juxtaposed with April 2020 deviations, model performance errors, and time-to-recovery. We identified 33 conditions that experienced the most intense utilization reductions in the post-hiatus phase, defined as having an average monthly decrease in utilization of 20% or more over this period that was statistically significant (i.e., δ_P_H ≤ -0.20 with p ≤ 0.05; also represented by dark shades of red in the circular heatmap of Figure 6); many of these conditions were asymptomatic, slowly progressing, and/or NVT (Table S5A). The diagnosis categories most represented in this set of conditions with the largest post-hiatus utilization reductions were cornea and external diseases (e.g., conjunctivitis-related diagnoses, peripheral corneal degeneration), followed by retinal and vitreous conditions (e.g., retinal microaneurysms, unspecified background retinopathy, venous engorgement, and “other retinal microvascular abnormalities”), oculofacial plastics and orbital conditions (e.g., in situ carcinoma of the eye, benign eyelid neoplasm, orbital floor fracture, and “other eyelid degenerative disorders”), and blindness and vision defects (e.g., visual loss, suspect amblyopia, and color vision deficiencies) (Table S5A). Conjunctivitis-related diagnoses were particularly well-represented among the set of conditions that exhibited intense post-hiatus utilization reductions, with presentations for infectious keratoconjunctivitis decreasing the most (δ_P_H = -0.38, 95% CI: -0.41 to -0.35, p < 0.001) among all diagnosis entities. No conditions that had a mean post-hiatus utilization reduction of 20% or more also recovered partially or fully, except for the diagnosis of eyelid/periocular superficial injury (δ_P_H = -0.21, 95% CI: -0.24 to -0.17, p < 0.001), which experienced a partial recovery in November 2020.

Few conditions (15/261 = 6%) met or exceeded counterfactual utilization predictions in the post-hiatus period (i.e., δ_P_H ≥ 0) (Table S5B); but among those that did, many were retinal and/or pediatric diseases, like unspecified DR with (δ_P_H = 0.46, 95% CI: 0.37 to 0.55, p < 0.001) and without (δ_P_H = 0.04, 95% CI: -0.01 to 0.09, p = 0.11) diabetic macular edema, infantile/juvenile cataract (δ_P_H = 0.17, 95% CI: 0.12 to 0.21, p < 0.001), eye injuries such as corrosion of the cornea/conjunctival sac (δ_P_H = 0.14, 95% CI: 0.08 to 0.21, p < 0.001) and ocular laceration without prolapse (δ_P_H = 0.09, 95% CI: 0.04 to 0.14, p < 0.001), and various stages of retinopathy of prematurity (ROP): ROP stage 3 (δ_P_H = 0.12, 95% CI: 0.06 to 0.18, p < 0.001), ROP stage 2 (δ_P_H = 0.04, 95% CI: -0.02 to 0.11, p = 0.17), and ROP with unspecified stage (δ_P_H = 0.05, 95% CI: 0.00 to 0.10, p = 0.07) All 15 diagnosis entities that met or exceeded post-hiatus counterfactual utilization levels also experienced recovery, with most of these conditions (12/15 = 80%) fully recovering.

Among all diagnosis entities, a broader set of conditions(116/261 = 44%) experienced some form of recovery (Table S6); however, many of these recoveries were not sustained (66/116 = 57%).The diagnosis categories with the highest proportions of conditions that experienced recoveries in utilization were uveitis and ocular inflammation (12/15 = 80%), post-operative complications (4/5 = 80%), ocular globe injuries/intraocular foreign bodies (3/4 = 75%), and cataract and other lens disorders (5/7 = 71%) (Figure 6). On the other hand, the diagnosis categories with the lowest proportions of conditions that experienced recovery were “other specified eye disorders" (2/13 = 15.4%), followed by refractive error (2/7 = 29%), retinal and vitreous conditions (28/75 = 37%), blindness and vision defects (5/13 = 38%), and cornea and external disease (22/56 = 39%) (Figure 6). Approximately half of all conditions in the diagnosis categories of oculofacial plastics and orbital conditions (15/29 = 51.7%), neuro-ophthalmology (8/16 = 50%), strabismus (3/6 = 50%), and glaucoma (7/15 = 47%) experienced recovery (Figure 6). . Among all diagnosis entities that experienced partial or full recovery, the most common month at which recovery occurred was June 2020 (42/116 = 36.2%), followed by September 2020 (17/116 = 14.7%), December 2020 (14/116 = 12.1%), June 2021 (14/116 = 12.1%), and February 2021 (13/116 = 11.2%) (Figure 6).

Monthly deviations for all diagnoses, along with 95% CIs and p-values (unadjusted and adjusted), are provided in the supplement (Figure S7, Tables S7-S9).

The prioritization of care for more severe conditions in the early pandemic phase has been previously reported, across other specialties³³, and among select ophthalmology practices.^34–36 However, the utilization patterns described in our national study not only encompass a wider spectrum of granular conditions, include more pandemic subperiods for analysis, and reflect comparisons from counterfactual utilization levels estimated via predictive models trained on multiple years of pre-pandemic data, but also reveal relationships between disease attributes and utilization. In doing so, we fill an area of needed research identified by a systematic review of pandemic utilization studies.³³ We observed an inverse relationship between a condition’s severity and its magnitude of underutilization during the pandemic, but there was heterogeneity in the strength and endurance of this relationship across different sets of diagnosis entities. This invites investigation into other factors to explain why some conditions may have been prioritized over others. For instance, exploring associations between pandemic utilization levels and other attributes of diagnosis entities, such as the degree of required follow-up clinical visits and treatment, could provide additional insights. Furthermore, the lack of a clear separation between the post-hiatus deviations of many diagnoses in our study could reflect a general reduction in the willingness of patient populations to seek care (due to, e.g., heightened patient risk perceptions, economic hardships, or shifts in habits) rather than resource limitations that constrained the provision of care. Older adults, which comprise a substantial proportion of the patient population seen in ophthalmology, may have been particularly susceptible to the avoidance of medical care. In a 2021 survey of 18,000 older adults from 11 high-income countries, American seniors were found to be most likely to experience economic difficulties related to the pandemic; and among seniors with two or more chronic conditions, those in the US reported postponements or cancellations of appointments most frequently.³⁷ Similarly, in the Netherlands, older adults with multiple chronic conditions were also more likely to avoid medical care.³⁸ Furthermore, older Black and Latino/Hispanic American adults were found to be substantially more likely to experience economic hardships than older white American adults.³⁷ These findings highlight the need to better understand the complex effects of the pandemic on the healthcare utilization patterns of vulnerable populations, and to develop targeted strategies for equitable access to care.

We also identified clusters of diagnoses that shared similar longitudinal utilization trends, thereby highlighting conditions that had above- and below-average utilization and distinguishing the relatively few conditions that experienced recovery from those that did not. Conditions exhibiting intense reductions of post-hiatus utilization mostly included those that were non-vision-threatening (e.g., conjunctivitis) or precursors of vision-threatening conditions (e.g., subclinical indications of DR like retinal microaneurysms/background retinopathy), whereas those that had above-average utilization included ocular emergencies with a risk for irreversible vision loss such as ROP. Further research, such as customized longitudinal cohort studies designed to investigate the long-term clinical impact of missed or delayed care for specific condition(s) of focus, is needed to understand the distinction between benign reductions in visits from potentially harmful ones like decreases in screenings for common conditions that can lead to irreversible vision loss.³⁹ Additionally, developing condition-specific definitions of loss to follow-up (LTFU) based on clinical considerations such as disease pathology and the appropriate frequency of follow-up or treatment,⁴⁰ enables targeted cohort studies to shed light on the patient-level determinants of LTFU during the COVID-19 pandemic.

This study has several limitations. Our parametric approach to establish counterfactual expectations of utilization introduces the possibility for model mis-specification. Furthermore, only three years’ worth of pre-pandemic data were used to train models because we restricted the analysis to only use one diagnostic coding system (ICD-10). However, model selection from a variety of candidate algorithms using leave-out-one-year cross validation enhanced the predictive ability of our counterfactual models, and even with three years of data, we were able to achieve low RMSPE. In some cases, observed deviations from expectation may be reflective of artificial factors that were extraneous to true changes in utilization, like adjustments in ICD-10 documentation practices and undetected data latency issues. Because we tried to account for data latency by excluding practices that did not contiguously contribute data throughout the study period, our analysis is biased towards practices that were able to remain operational throughout the pandemic, despite sensitivity analyses demonstrating no major changes in primary outcomes. Additionally, due to current database limitations and the large amount of diagnosis entities included for analysis, we were unable to directly identify the primary condition or disease at each patient visit, and could not readily distinguish, at scale, newly occurring incident diagnoses from ones that may be repeat documentations of historical conditions or recurrent/ongoing clinical problems. Although this study used the largest and most comprehensive registry of eye diseases in the United States, our results may not capture potential shifts in ocular care volume that occurred during the pandemic from practices integrated with the IRIS Registry to eye care settings not covered by the database (e.g., at optometrists who are not employed by ophthalmologists, or tertiary academic medical centers that are not included in the IRIS Registry). Finally, there are inherent ambiguities associated with terms like “severity”, which are also reflected in the external measures of disease severity we used for hypothesis testing.

Nonetheless, given the COVID-19 pandemic’s lasting effects on the allocation of healthcare resources, the public health implications of widespread and prolonged disruptions to care are increasingly important to understand.¹⁰ An exploratory but expansive characterization of care utilization patterns across a specialty may serve as an initial tool for researchers and clinicians to evaluate the extent, magnitude, and differential impact of these disruptions. Here, using data from the United States’ first comprehensive eye disease registry, we employed a common analytical framework to identify patterns of patient presentations during the first two years of the COVID-19 pandemic across a broad set of granular ophthalmic conditions. This framework can be flexibly adapted to examine longitudinal pandemic utilization trends among services from other specialties and across different healthcare settings. High-dimensional and high-resolution characterizations of care utilization may enable timely and ongoing assessments of the overall state of healthcare usage across a specialty, inform targeted efforts to investigate the pandemic’s long-term impact on clinical outcomes and burdens of disease, and highlight factors that may be driving the differential utilization of care.

This study did not require IRB/ethics committee approval because the data used are from a deidentified clinical registry, This study was approved and overseen by the Quality and Data Science Division of the American Academy of Ophthalmology and adhered to the tenets of the Declaration of Helsinki.

DATA AVAILABILITY

The original data that support the findings of this study are from the American Academy of Ophthalmology IRIS® Registry, which is not a publicly available dataset. However, the minimum data necessary to interpret, verify, and extend the research are currently available in the Supplementary Information and at https://github.com/charlesli37/covid-oph-dx-utils. Eligible investigators may apply for research opportunities to work with IRIS Registry data; more information is available at https://www.aao.org/iris-registry/data-analysis/requirements or by contacting [email protected].

CODE AVAILABILITY

SQL and R programming scripts used for this study are available at https://github.com/charlesli37/covid-oph-dx-utils.

World Health Organization. Third round of the global pulse survey on continuity of essential health services during the COVID-19 pandemic: November–December 2021. 2022.
McGough M, Amin K, Cox C. How has healthcare utilization changed since the pandemic? [Internet]. Peterson-KFF Health System Tracker. 2023 [cited 2023 Apr 8];Available from: https://www.healthsystemtracker.org/chart-collection/how-has-healthcare-utilization-changed-since-the-pandemic/
Mehrotra A, Chernew M, Linetsky D, Hatch H, Cutler D, Schneider EC. The impact of COVID-19 on outpatient visits in 2020: Visits remained stable, despite a late surge in cases [Internet]. The Commonwealth Fund. 2021;Available from: https://www.commonwealthfund.org/publications/2021/feb/impact-covid-19-outpatient-visits-2020-visits-stable-despite-late-surge
Gillis K. Impacts of the COVID-19 Pandemic on 2020 Medicare Physician Spending [Internet]. American Medical Association; 2021. Available from: https://www.ama-assn.org/system/files/2020-prp-covid-impact-medicare-physician-spending.pdf
Gallagher K, Gerhart J, Amin K, Rae M, Cox C. Early 2021 data show no rebound in health care utilization [Internet]. Peterson-KFF Health System Tracker. 2021 [cited 2022 Mar 29];Available from: https://www.healthsystemtracker.org/brief/early-2021-data-show-no-rebound-in-health-care-utilization/
Mehrotra A, Chernew M, Linetsky D, Hatch H, Cutler D. What impact has COVID-19 had on outpatient visits? [Internet]. The Commonwealth Fund. 2020;Available from: https://www.commonwealthfund.org/publications/2020/apr/impact-covid-19-outpatient-visits
McKinsey & Co. COVID-19 Hospital Insights Survey [Internet]. McKinsey & Company; February, 18 2022 [cited 2022 Mar 29]. Available from: https://www.mckinsey.com/industries/healthcare-systems-and-services/our-insights/survey-us-hospital-patient-volumes-move-back-towards-2019-levels
Cheung SSL, Wong CYK, Chan JCK, et al. Ophthalmology in the time of COVID-19: experience from Hong Kong Eye Hospital. Int J Ophthalmol 2020;13(6):851–9.
Chen EM, Parikh R. COVID-19 and ophthalmology: The pandemic’s impact on private practices. EyeNet 2020;
Cutler DM. How COVID-19 Changes the Economics of Health Care. JAMA Health Forum 2021;2(9):e213309–e213309.
Ellis RP, Martins B, Zhu W. Health care demand elasticities by type of service. J Health Econ 2017;55:232–43.
Zhou Z, Su Y, Gao J, Xu L, Zhang Y. New estimates of elasticity of demand for healthcare in rural China. Health Policy 2011;103(2–3):255–65.
Khan JA, Mahumud RA. Is healthcare a “Necessity” or “Luxury”? an empirical evidence from public and private sector analyses of South-East Asian countries? Health Econ Rev 2015;5:3.
Mafi JN, Craff M, Vangala S, et al. Trends in US Ambulatory Care Patterns During the COVID-19 Pandemic, 2019-2021. JAMA 2022;327(3):237–47.
Whaley CM, Pera MF, Cantor J, et al. Changes in Health Services Use Among Commercially Insured US Populations During the COVID-19 Pandemic. JAMA Netw Open 2020;3(11):e2024984.
Czeisler MÉ, Marynak K, Clarke KEN, et al. Delay or Avoidance of Medical Care Because of COVID-19-Related Concerns - United States, June 2020. MMWR Morb Mortal Wkly Rep 2020;69(36):1250–7.
Radhakrishnan L, Carey K, Hartnett KP, et al. Pediatric Emergency Department Visits Before and During the COVID-19 Pandemic - United States, January 2019-January 2022. MMWR Morb Mortal Wkly Rep 2022;71(8):313–8.
Agency for Healthcare Research and Quality. Clinical Classifications Software Refined (CCSR) for ICD-10-CM Diagnoses [Internet]. Healthcare Cost and Utilization Project; 2020. Available from: https://www.hcup-us.ahrq.gov/toolssoftware/ccsr/dxccsr.jsp
American Academy of Ophthalmology. EHR Systems - IRIS® Registry [Internet]. American Academy of Ophthalmology. 2023 [cited 2023 Apr 11];Available from: https://www.aao.org/iris-registry/ehr-systems
Chiang MF, Sommer A, Rich WL, Lum F, Parke DW 2nd. The 2016 American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight) Database: Characteristics and Methods. Ophthalmology 2018;125(8):1143–8.
Bergmeir C, Benítez JM. On the use of cross-validation for time series predictor evaluation. Inf Sci 2012;191:192–213.
Roberts DR, Bahn V, Ciuti S, et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 2017;40(8):913–29.
Kettenring JR. The Practice of Cluster Analysis. J Classification 2006;23(1):3–30.
Wilkinson L, Friendly M. The history of the cluster heat map. Am Stat 2009;63(2):179–84.
Gu, Eils, Schlesner. Complex heatmaps reveal patterns and correlations in multidimensional genomic data. Bioinformatics.
Gu, Gu, Eils, Schlesner, Brors. circlize implements and enhances circular visualization in R. Bioinformatics.
Bourges J-L, Boutron I, Monnet D, Brézin AP. Consensus on Severity for Ocular Emergency: The BAsic SEverity Score for Common OculaR Emergencies [BaSe SCOrE]. J Ophthalmol 2015;2015:576983.
U.S. Centers for Disease Control and Prevention. Common Eye Disorders and Diseases [Internet]. CDC Vision Health Initiative (VHI). 2020 [cited 2022 Jun 21];Available from: https://www.cdc.gov/visionhealth/basics/ced/index.html
Benjamini Y, Hochberg Y. Controlling the false discovery rate: A practical and powerful approach to multiple testing. J R Stat Soc 1995;57(1):289–300.
Wickham H. ggplot2: Elegant Graphics for Data Analysis [Internet]. 2016;Available from: https://ggplot2.tidyverse.org
Amazon Web Services. Amazon Redshift Documentation [Internet]. 2022. Available from: https://docs.aws.amazon.com/redshift/?id=docs_gateway
R Core Team. R: A Language and Environment for Statistical Computing [Internet]. 2021;Available from: https://www.R-project.org/
Moynihan R, Sanders S, Michaleff ZA, et al. Impact of COVID-19 pandemic on utilisation of healthcare services: a systematic review. BMJ Open 2021;11(3):e045343.
Berkenstock MK, Liberman P, McDonnell PJ, Chaon BC. Changes in patient visits and diagnoses in a large academic center during the COVID-19 pandemic. BMC Ophthalmol 2021;21(1):139.
Al-Khersan H, Kalavar MA, Tanenbaum R, et al. Emergent Ophthalmic Surgical Care at a Tertiary Referral Center During the COVID-19 Pandemic. Am J Ophthalmol 2021;222:368–72.
Breazzano MP, Nair AA, Fernando Arevalo J, et al. Frequency of Urgent or Emergent Vitreoretinal Surgical Procedures in the United States During the COVID-19 Pandemic. JAMA Ophthalmol 2021;139(4):456–63.
Williams RD II, Shah A, Doty MM, Fields K, FitzGerald M. The Impact of COVID-19 on Older Adults [Internet]. The Commonwealth Fund. 2021 [cited 2023 Apr 12];Available from: https://www.commonwealthfund.org/publications/surveys/2021/sep/impact-covid-19-older-adults
Schuster NA, de Breij S, Schaap LA, et al. Older adults report cancellation or avoidance of medical care during the COVID-19 pandemic: results from the Longitudinal Aging Study Amsterdam. Eur Geriatr Med 2021;12(5):1075–83.
Berwick DM, Hackbarth AD. Eliminating waste in US health care. JAMA 2012;307(14):1513–6.
Khurana RN, Li C, Lum F. Loss to Follow-up in Patients with Neovascular Age-Related Macular Degeneration Treated with Anti-VEGF Therapy in the United States in the IRIS® Registry. Ophthalmology 2023;130(7):672–83.

There is NO Competing Interest.

COVIDElasticityPaperSupplementCommsMedRevised.pdf
Supplementary Information
nrreportingsummaryRevised.pdf
Reporting Summary

Download PDF

Journal Publication

published 14 Dec, 2023

Read the published version in Communications Medicine →

Version 1

posted

You are reading this latest preprint version

Shifts in Care Utilization Patterns During the COVID-19 Pandemic: A High-Dimensional Study of Presentations for Ophthalmic Conditions in the US

Status:

Journal Publication

Version 1

Abstract

Figures

INTRODUCTION

METHODS

RESULTS

DISCUSSION

DECLARATIONS

REFERENCES

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1