Disrupted variance in respiratory illness presentation patterns in New Zealand primary care through and after the SARS-CoV-2 pandemic: an observational cohort study using automated mining of electronic health records and clinical text

Abstract

Objectives

To identify and compare respiratory illness presentation rates and service utilisation to primary care in Aotearoa, New Zealand (NZ) before, through and after the COVID-19 pandemic.

Design

A natural language processing (NLP) software interference algorithm was used to interrogate quantitative and qualitative cross-sectional and retrospective cohort data from clinical consultation notes.

Setting

Primary care presentations across Aotearoa, NZ, encompassing both urban and rural areas.

Participants

Electronic records were obtained from general practice interactions from 2018 to end of 2023. Participants consisted of 37 consenting practices from eight districts.

Results

A total of 10 421 399 unique medical interactions were processed by the NLP software, of which 15.1% (1 573 631) were related to a respiratory issue. This ranged from 14.2% and 13.7% in 2018 and 2019 to a peak of 25.2% in 2022. Expected patterns of seasonal variation were observed prepandemic with small variance in mean rates. Due to the COVID-19 pandemic, border closures and then SARS-CoV-2 arrival, there was significant disruption to previous seasonal patterns across all respiratory illness presentations in primary care. Patterns vary markedly across different respiratory groups. While the degree of variance has reduced, there continues to be marked disruption of pre-COVID patterns.

Conclusions

Primary care respiratory presentations in NZ continue to experience persistent variation from the pre-COVID typical seasonal presentation patterns. There is significant change in presentations, likely due to changing disease prevalence, but also to changing patient and provider behaviour. These results have implications for primary care workload management and public health policy.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study uses a validated natural language programming algorithm to explore respiratory illness presentation and utilisation patterns to primary care, before, during and post the COVID-19 pandemic. A total of 10 421 399 unique medical interactions were assessed.

• Due to the extensive time period reviewed, the results are a significant addition to information about the disruption caused by COVID-19 to previously well-established seasonal variation in different respiratory illness categories. This has important implications for clinical practice and workforce planning.

• The algorithm assesses patient presentation for different respiratory illness categories, providing insights into primary care utilisation and workload, rather than presumed ‘disease incidence’. A COVID-19 presentation, for example, is an important element of workload, even when a person is COVID-19 negative.

• Information about the consultation is based on doctor decision making, and the algorithm is limited by the quantity and detail of the recorded information in each consultation in the electronic health record.

• The exclusion of specific out-of-hours consultations may result in an underestimation of primary care respiratory tract-related presentation rates.

Background

Internationally, respiratory illness, both acute and chronic, represents a substantial proportion of primary care workload.^{1 2} In Aotearoa, New Zealand (NZ), upper respiratory illness is one of the most common diagnoses³ and chronic conditions, such as asthma and chronic obstructive pulmonary disease (COPD), alongside acute respiratory infections posing significant public health impact.

Understanding the pattern and burden of respiratory illness in primary care is essential for maintaining effective health systems, improving public health and optimising healthcare resources.

Healthcare systems worldwide experienced substantial workload challenges during the COVID-19 pandemic, affecting both hospital and primary care services,^{4 5} with primary care playing a crucial role in pandemic response.^{6 7}

Recent patient cohort studies have assessed changes in primary care presentations in the early stages of the COVID-19 pandemic. A large-scale international consortium analysis found marked decreases in primary care visits for respiratory illness throughout the pandemic.⁸ While seasonal variations in allergic respiratory illnesses persisted across many countries, influenza and other respiratory infections did not experience typical seasonal spikes in the first 2 years of the pandemic. Despite a reduction in respiratory visits, primary care continued to provide services for chronic respiratory condition management. Internationally, there is limited research on the long-term postpandemic effects on these healthcare patterns.

The NZ strict response to COVID-19 in March 2020 involved border closures and early lockdowns. These measures led to a very reduced circulation of SARS-CoV-2 compared with many countries, with resulting low levels of COVID-19 disease prior to the introduction of the COVID-19 vaccination in February 2021. The country achieved very high levels of vaccination uptake during the initial phases of the pandemic, with 90% of eligible population being vaccinated in late 2021 and 95% by early 2022. The border closures and lockdowns also disrupted other viral circulation patterns. This included absent Respiratory Syncytial Virus (RSV) community transmission for 1 year and influenza for 2 years leading to a unique opportunity to observe utilisation patterns of respiratory illness.^{9 10} There were partial border relaxations in April 2021 and in February 2022.

Respiratory syncytial virus

community transmission for 1 year and influenza for 2 years leading to a unique opportunity to observe utilisation patterns of respiratory illness.^{9 10} There were partial border relaxations in April 2021 and in February 2022.

A longstanding challenge in healthcare and particularly primary care is accessing the rich but unstructured information about patient care and presentation contained in electronic health record (EHR) systems. While secondary care services have had relatively straightforward access to respiratory admission and discharge data, much less information is known about presentation patterns in primary care.⁸ Despite increased coding use in general practice, usually in research network settings, overall presentation coding remains limited, requiring additional techniques to try and ‘unlock’ patient note data.

The last 10 years have seen the potential for ‘big data’ assessment of general practice workload. Using advancements in computational resources, text classification alongside machine learning algorithms, extraction of clinical narratives from unstructured medical notes can lead to effective and efficient analysis of large data sets in primary care.¹¹

In NZ, natural language processing (NLP) has been used previously to explore presentation patterns in influenza-like illness, childhood respiratory illness and other medical conditions.^{3 11 12} There has been a successful use of NLP approaches to identify COVID-19 presentations to general practice in the Netherlands.¹³

Understanding primary care workload patterns and presentation will help support strategies to mitigate workload pressures when planning for future pandemics. Some existing studies show, for example, that healthcare workers, particularly those in frontline roles, were especially vulnerable to burnout,¹⁴ with higher rates reported across all specialties postpandemic.15,17

Given the NZ particular COVID-19 management environment, and our NLP capability, we aimed to analyse primary care utilisation data and observe respiratory illness patterns presenting to primary care before, during and for a substantial period after the COVID-19 pandemic.

Methods

In NZ, patients sign consent forms on general practitioner (GP) enrolment agreeing to anonymised data use in ethically approved research. Data use for this study is in accord with this agreement, and anonymised data from EHRs from recruited general practices was available for analysis.

A natural language processing (NLP) software interference algorithm was used to interrogate quantitative and qualitative cross-sectional and retrospective cohort data from EHR.^{11 12} Interactions were classified using a novel text classifier categorising interactions based on recorded clinical features. Data were collected retrospectively to cover the period before, during and after the pandemic.

General practice populations were chosen to specifically reflect there was appropriate representation for Māori, Pacific and the most socio-economically deprived groups given their known risk of poorer respiratory health outcomes.

In NZ, every person who receives healthcare has a National Health Index (NHI) number. This is a unique identifier assigned to each person. The NHI number allows for tracking healthcare interactions with both primary and secondary care throughout NZ. It also links to other data in the EHR, including ethnicity and socio-economic status. Socioeconomic deprivation is measured using the NZ Index of Socioeconomic Deprivation, which is a small-area measure based on census data and assigned to deciles, with decile 1 representing the least deprived and decile 10 representing the most deprived areas. We have aggregated into quintiles, with 1 being least and 5 the most deprived.

Process

The software inference algorithm analysed electronic medical records related to all interactions in the selected time period. It categorised each interaction based on the clinical information documented by GPs and practice nurses, including free-text clinical narrative, diagnostic classifications, using read codes and prescribed medications. All interactions were included irrespective of mode (eg, face-to-face, telephone, secure message).^{3 12}

Each interaction was assessed against multiple respiratory classes. These categories were influenza-like illness (ILI), lower respiratory tract illness (LRTI), non-URTI rhinitis (NUR), otitis media, sore throat, upper respiratory tract infection (URTI) and wheeze illness, COVID-19 positive, COVID-19 negative or not otherwise specified. Interactions were further assigned classes for respiratory, non-COVID-19 respiratory, asthma and other chronic lung disease, which includes chronic obstructive pulmonary disease (COPD) using appropriate demographic and clinical rulesets.

To differentiate between wheeze, asthma and chronic lung disease, the algorithm was designed to classify these conditions as a single class, and a subclassification was applied through a series of rules according to age and smoking status. Patients under 5 years were classified as ‘wheeze illness’, if aged over 5 years, classified as asthma, unless they had a COPD code, another chronic lung disease code or were smokers over the age of 50.

Interactions consisting of standard respiratory screening activity with no positive findings or respiratory signs or symptoms were excluded to avoid overestimation of respiratory tract illness burden.

COVID-19 presentations were classified in two ways; interactions where there was a positive diagnosis of COVID-19 (COVID_ POS), and those where the interaction concerned COVID-19, but the result was either negative, unspecified or there was no evidence of a test.

ILI was classified using two different clinical definitions. The first was using the WHO definition of an acute respiratory infection with onset within the past 10 days, presenting with cough and a history of a fever or a measured temperature of ≥38°C. The second was using a definition which was consistent with prior research, which matched those under 5 with any upper respiratory tract sign or symptom, or those over 5 with a temperature of 38 degrees or more with one or more upper respiratory tract signs or symptoms. We included the later class to maintain consistency with a prior study of ILI presentation from the H1N1 pandemic of 2009¹⁴ and to represent a clinically appropriate general practice presentation of ILI.

The algorithm³ underwent training, testing and validation using gold-standard datasets, each containing 1200 consultation records that were independently classified by two general practice clinical experts. The algorithm was designed to replicate the judgements made by these clinical experts. Development aimed to optimise specificity while maximising sensitivity to minimise the occurrence of false positives. The algorithm was trained using a fitness metric of the harmonic mean of model precision and recall (F1 score). The algorithm’s sensitivity, specificity, positive and negative predictive values, and F1 measure for each of the diagnostic categories was matched against the gold standard validation set of 1200 consultation records.

A workflow diagram is provided as a supplementary figure (online supplemental figure 1) to describe in more detail how the EHR data is processed, classified and analysed.

Patient and public involvement statement

In NZ, patients sign consent forms on GP enrolment agreeing to anonymised data use in ethically approved research. Data use for this study is in accord with this agreement, and hence patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

Results

Setting and participants

All individuals in NZ are enrolled with a primary care practice and have a unique NHI number enabling identification of interactions across time and location, and linkage with other information in the EHR such as ethnicity and socio-economic status.³ The study consisted of 37 consenting practices from eight districts across NZ. The study cohort consists of 467 736 individuals and was compared with the 2018 census population of 4 699 755 individuals.

Table 1 compares the study cohort with the census population. Age-specific rates were calculated and applied to the NZ standard population using the 2018 census, to adjust for age differences. The age categories are divided into 5-year age bands. Individuals without known ages were excluded from this calculation.

Table 1. Participant characteristics: percentage of study population compared with census 2018 data.

	Study (n=467 736)	Census (n=4 699 755)	Difference
Ethnicity
Asian	16.0	14.3	1.7
European	53.5	62.6	−9.1
Māori	14.3	15.3	−1.0
Other	8.7	2.3	6.4
Pacific	7.5	5.5	2.0
Gender
Male	48.9	49.4	−0.5
Female	50.5	50.6	−0.1
Other	0.5		0.5
Age
0–4 years	10.3	6.3	4.0
5–9 years	4.9	6.9	−2.0
10–14 years	5.1	6.5	−1.4
15–19 years	6.0	6.4	−0.4
20–24 years	7.8	6.8	1.0
25–29 years	9.4	7.3	2.1
30–34 years	8.7	6.7	2.0
35–39 years	7.1	6.3	0.8
40–44 years	6.2	6.2	0.0
45–49 years	6.3	6.8	−0.5
50–54 years	6.0	6.6	−0.6
55–59 years	5.6	6.4	−0.8
60–64 years	4.8	5.6	−0.8
65–69 years	4.0	4.9	−0.9
70–74 years	2.9	3.9	−1.0
75–79 years	2.1	2.8	−0.7
80–84 years	1.3	1.8	−0.5
85 years and over	1.5	1.8	−0.3
Deprivation quintile
0 (unknown)	5.8	0.1	5.7
1 (least)	18.3	20.6	−2.3
2	18.7	20.6	−1.9
3	18.7	20.5	−1.8
4	19.0	20.1	−1.1
5 (most)	19.5	18.0	1.5

To explore generalisation and validity of the setting and participants, comparison has been made as shown with age bands, ethnicity, gender and socio-economic deprivation quintiles.

Performance of the software inference algorithm

Table 2 shows the performance of the algorithm, with Matthews Correlation Coefficient (MCC) as the primary fitness metric with 95% CIs. The following classes are composite classes:

Table 2. Software inference algorithm performance.

Class	mcc	mcc_ci	ppv	ppv_ci
ASTHMA_COPD_WHEEZE	0.896	(0.872 to 0.921)	0.883	(0.852 to 0.921)
ANY	0.868	(0.849 to 0.884)	0.887	(0.865 to 0.908)
NOT RESPIRATORY	0.865	(0.846 to 0.882)	0.980	(0.976 to 0.984)
ILI_INT	0.828	(0.748 to 0.904)	0.818	(0.706 to 0.909)
SORE_THROAT	0.824	(0.774 to 0.867)	0.807	(0.745 to 0.870)
NUR	0.821	(0.768 to 0.874)	0.804	(0.739 to 0.878)
ILI_Legacy	0.803	(0.748 to 0.853)	0.738	(0.667 to 0.810)
COVID	0.797	(0.756 to 0.841)	0.786	(0.733 to 0.843)
RTI	0.755	(0.714 to 0.795)	0.722	(0.675 to 0.771)
URTI	0.733	(0.684 to 0.775)	0.674	(0.609 to 0.730)
LRTI	0.713	(0.621 to 0.792)	0.750	(0.654 to 0.867)
COVID_NOS	0.686	(0.621 to 0.744)	0.667	(0.596 to 0.739)
COVID_POS	0.664	(0.546 to 0.763)	0.679	(0.550 to 0.800)
OTITIS_MEDIA	0.522	(0.299 to 0.706)	0.750	(0.500 to 1.000)

Class	sensitivity	sensitivity_ci	specificity	specificity_ci
NOT RESPIRATORY	0.979	(0.975 to 0.983)	0.889	(0.868 to 0.908)
ASTHMA_COPD_WHEEZE	0.918	(0.894 to 0.950)	0.992	(0.990 to 0.995)
ANY	0.889	(0.869 to 0.909)	0.980	(0.975 to 0.984)
ILI_Legacy	0.878	(0.826 to 0.936)	0.993	(0.991 to 0.995)
NUR	0.846	(0.780 to 0.917)	0.996	(0.994 to 0.997)
SORE_THROAT	0.845	(0.787 to 0.904)	0.995	(0.993 to 0.997)
ILI_INT	0.842	(0.750 to 0.947)	0.998	(0.997 to 0.999)
COVID	0.827	(0.773 to 0.877)	0.991	(0.989 to 0.994)
URTI	0.822	(0.768 to 0.875)	0.986	(0.983 to 0.989)
RTI	0.815	(0.770 to 0.863)	0.984	(0.981 to 0.988)
COVID_NOS	0.725	(0.647 to 0.800)	0.991	(0.988 to 0.993)
LRTI	0.679	(0.571 to 0.781)	0.997	(0.995 to 0.998)
COVID_POS	0.655	(0.524 to 0.783)	0.996	(0.995 to 0.998)
OTITIS_MEDIA	0.364	(0.167 to 0.556)	1.000	(0.999 to 1.000)

ANY: made up of all other classes, positive when any other class is positive.

COVID: made up of COVID_NOS and COVID_POS, positive when either is positive.

RTI: made up of URTI and LRTI, positive when either is positive.

The worst performing class was OTITIS_MEDIA, with a MCC of only 0.522 and sensitivity of only 0.364. The algorithm was only able to detect just over one in three presentations of otitis media. This results in difficulty in inferring much about otitis media presentations in our analysis, particularly when any subcategory analysis is relied on.

All other categories performed moderately well across accuracy metrics.

A total of 10 421 399 unique medical interactions were processed, of which 15.1% (1 573 631) were related to a respiratory issue. This ranged from 14.2% and 13.7% in 2018 and 2019 to a peak of 25.2% in 2022. The respiratory interaction rate was analysed for 2018 (14.2%), 2019 (13.7%), 2020 (15.5%), 2021 (16.0%), 2022 (25.2%) and 2023 (15.3%).

The following figures show the changes in general practice presentation over time, with clear differentiation between the pre-COVID patterns of seasonality, the disruption caused by COVID-19 border restrictions and their aftermath. NZ operated under a series of Protection Framework levels. In level 1, there were no restrictions; in level 2, there were limits on gatherings; in level 3, only purposeful travel was allowed and there were strict limits on gatherings; and in level 4, only essential travel was allowed and gatherings were banned. The figures allow comparison of primary care interactions with these protection levels (figure 1).

Figure 1. Age-adjusted respiratory interaction rates and all interaction counts for the cohort primary care practices. NZ, New Zealand; PF, Protection Framework.

Prior to COVID-19, in 2018 and 2019 there is a clear seasonal pattern in total interactions. This was disrupted in 2020 and 2021 with less apparent seasonal variation, with a lower but wider distribution. 2022 exhibited more overall interactions with a peak in summer 2022 and then a decline in 2023, though presentations remained high.

Figure 2 shows COVID-19 related interactions compared with overall respiratory illness presentation. There is a distinct seasonal pattern for overall respiratory presentations (ANY) in 2018 and 2019, then an out-of-season spike in all respiratory presentations associated with March 2020. This increase in interactions in March 2020 is attributed to a sharp spike in respiratory presentations which are not related to COVID-19 positive presentations as there was no actual circulating COVID-19 disease at the time and the country was in ‘lockdown’. There was, at this time, a spike specifically of COVID-19 related (negative or not specified) presentations as the population presented with concerns about possible COVID-19 incursions.

Figure 2. Unique interactions with a primary care clinical by month for all respiratory illness, compared with presentations for COVID-19 illness, COVID-19 related presentations, COVID-19 positive presentations and all respiratory presentations that were not COVID-19. NZ, New Zealand; PF, Protection Framework.

Throughout 2020, there was a lack of the usual winter seasonal peak in non-COVID respiratory presentations. The non-COVID-19 respiratory presentations showed the pre-COVID seasonal peak again from 2021 onwards. The non-COVID-19 respiratory presentations were overall slightly lower post-COVID than pre-COVID.

There was a significant spike in COVID-19 positive presentations in 2022. This coincides with the reduction in border restrictions and the arrival and circulation of COVID-19 delta strain and then omicron strains.

There continued to be fairly consistent presentations for COVID-19 related presentations (negative or not otherwise specified) following the arrival of COVID-19 in NZ in 2020. Since the winter of 2022, overall COVID presentations have reduced. This reduction has been greatest for COVID-19 positive presentations, with less decrease in those related to COVID-19 but not specifically COVID-19 positive.

Figure 3 shows the changing patterns of different respiratory interactions by the different categories. All categories except NUR show seasonal winter peaks pre-COVID-19. NUR shows seasonal, southern hemisphere summer peaks with allergic rhinitis presentations and prescriptions featuring prominently. Seasonal patterns all appear to have been disrupted in 2020 except NUR.

Figure 3. Unique interactions with a primary care clinical by month by type of respiratory illness presentation. NZ, New Zealand; PF, Protection Framework.

Asthma and chronic lung disease showed increased activity spikes in 2020 at the start of the first NZ lockdown, with asthma presentations in particular peaking at this time. The seasonal pattern for asthma returned by winter 2021. Post the COVID-19 main period, during winter 2022/2023 there were slightly higher peaks and troughs for asthma seasonal patterns.

Wheeze illness (children aged under 5 years) peaked sharply at the start of the first lockdown in 2020, disrupting the typical seasonal pattern and then followed by a very high winter peak in 2021. Following this, there was a return to seasonal peaks, but at a lower level than pre-COVID-19.

Chronic lung disease showed a significant spike at the start of lockdown in 2020, and while there was some disruption in 2020 and 2021, seasonal patterns have returned since 2022, although with lower peak levels.

Both classes of ILI show clear seasonal patterns in 2018/2019. In 2020 there was no standard winter seasonal peak, with rates similar to the usual season trough. There were increases in rates in 2021, but lower than usual for peaks. Both had a winter peak similar to pre-COVID in 2022 though with a magnitude lower than 2018/2019. Both classes of ILI have very similar patterns, although overall ILI International has an overall rate about half that of ILI Legacy.

For upper respiratory tract infections (URTIs), there was significant disruption to seasonal patterns in 2020 followed by a rapid return to seasonal patterns in 2021.

LRTIs followed seasonal pre-COVID trends but dropped at lockdown onset. There was no winter peak in 2020 with seasonal peaks returning from 2021 but with lower peaks. The seasonal trough in 2022 was much lower than normal, th

ough seasonal peaks have since been rising, with a higher peak in 2023 than 2022.

Otitis media presentations are represented by low numbers compared with other categories, but there appear to be lower rates since the COVID-19 pandemic.

Sore throat presentations showed disruption since the COVID-19 pandemic onset, with a more prolonged winter peak in 2021 lower peaks in 2023.

Figure 4 describes the overall age rates for respiratory illness presentation by age. There are clear pre-COVID-19 seasonal peaks with the highest rates consistently for younger children, 0–4 years, although in 2020 and 2021 rates for this group were much lower than pre-COVID-19. Through the COVID-19 period, the rates of presentation for younger children mostly stayed the highest; however, at the peak of the pandemic, rates for older children (5–18) and mid-life adults (19–64 years) increased relative to the other age groups. By 2023 and 2024, the patterns have returned to similar to pre-COVID, although younger children were still presenting at lower rates.

Figure 4. Unique interactions with a primary care clinical by month by age group for all respiratory illness presentations. PF: Protection Framework.

Discussion

By using automated mining of EHRs and clinical text, our study demonstrates the ongoing significant burden of respiratory illness work undertaken in primary care and shows profound changes in the prevalence, seasonality and patterns of respiratory illness presentation through the COVID-19 pandemic. While previous studies have described changes in respiratory tract consultations to primary care during the first part of the pandemic,^{18 19} our results explore changes in respiratory presentation and primary care utilisation through and beyond the pandemic, into the postpandemic period.

Our results should be seen within the context of the causal relationship between NZ’s pandemic response measures and changes in the observed patterns of respiratory care. Prior to the pandemic, NZ exhibited consistent annual seasonal variation in primary care presentations for respiratory illness with minimal deviations in mean presentation rates. From March 2020, NZ undertook border closures, quarantine, lockdowns, social distancing and other non-pharmaceutical measures to prevent the entry of SARS-CoV-2 into the country.⁹ This ‘elimination’ strategy enabled population protection prior to the introduction of the COVID-19 vaccination in February 2021 followed by high levels of vaccination uptake. The eventual reopening led to SARS-CoV-2 reintroduction and the significant disruption to established seasonal patterns across all illness categories. This unique situation also resulted in complete absence of influenza transmission for 2020²⁰ and 2021 and disruption of many other respiratory virus patterns including RSV.¹⁰

During the early pandemic, despite the absence of SARS-CoV-2, respiratory consultations remained high due to heightened COVID-19 concerns. We also note unexpected higher presentation rates for more chronic, usually non-infectious respiratory conditions such as asthma and other chronic lung disease. This combined mixed pattern resulted in overall presentation rates remaining high even when SARS-CoV-2 was not circulating and other viruses were at low circulating levels.

International comparison of our results must be viewed within the context of NZ COVID-19 policy and geography. The country response was characterised by early and very strong border controls, an initial elimination strategy and then domestic restrictions balanced with long border closures. The overall COVID result was aided by the geographical advantage of being an island nation with no land borders.

Changes in overall respiratory presentations during the pandemic have been noted internationally. A decline in emergency department visits for non-COVID-19 respiratory issues was noted for children in Israel, as parents/caregivers were discouraged or hesitant to present to hospitals due to lockdowns for fear of infection.²¹ Similar to our results, that study also noted changes in seasonality and lower rates of overall respiratory disease in comparison to prepandemic years, particularly for asthma, pneumonia and bronchiolitis.

Similar to our findings, Westfall et al noted a significant decline in primary care respiratory illness attendance across nine countries, due to widespread mitigation measures.⁸ In contrast, they noted less reduction in visits for chronic respiratory conditions, such as asthma.

Our results show an initial rise in asthma and other chronic respiratory conditions activity. As SARS-CoV-2 rates peaked in late 2021, higher rates of asthma in the following year may have been a follow-on effect from circulating COVID-19, or preparedness for needing medication through periods of both circulating virus and intermittent lockdowns and social restrictions which were finally fully lifted in late 2022.

Other studies also note a reduction in asthma activity during the pandemic.²² Given contrasting results, the impact of SARS-CoV-2 on asthma has been the source of considerable discussion.²³ In common with other studies, we suggest it is likely that behavioural factors and dynamic changes in perceptions of risk may have led to a change in presentation patterns.²⁴

Given asthma’s primary care burden, preparedness for fluctuating presentation rates should be incorporated into pandemic planning.²⁵

Our results show seasonal presentation patterns continued to be markedly disturbed for at least 3 years postpandemic onset. While variance from traditional seasonal patterns has since reduced with a more traditional winter peak returning by mid-2023, after 4 years postpandemic onset, there remains marked disturbance in pre-COVID-19 seasonal trends. Changes to seasonality have also been noted internationally; in Taiwan, alterations to prepandemic patterns at a population level continued for 3 years postpandemic.²⁶

We noted a relatively lower proportion of presentations for LRTIs at the start of the pandemic. This may have reflected initial pandemic changes in health-seeking behaviour; NZ promoted strong social distancing messaging particularly for older and high-risk groups, potentially shifting behaviours to manage respiratory illness at home. However, there continued to be lower peaks of LRTIs which may reflect initially the absence of transmission of SARS-CoV-2 in the NZ environment, followed by the absence of transmission of other viruses, particularly RSV and influenza.¹⁰ There is little comparable general practice information about LRTI presentations, though a German study also noted reduced incidence during COVID-19.²⁷

In contrast, milder upper respiratory tract infection rates showed less variation and returned earlier to prepandemic levels, likely reflecting the continued circulation of other common respiratory pathogens and the lesser disruption to these virus circulation patterns by non-pharmaceutical interventions.¹⁰

Shifts in primary care service delivery during the pandemic included rapid transitions from in-person to remote/online consultations to limit infection risks²⁸ with elderly patients particularly affected by decreased access to primary care services.²⁹ Our study did not show decreased presentations for older people, potentially a reflection on NZ general practice services’ ability to respond rapidly to a mixed online and face-to-face model.³⁰

While some seasonal patterns are slowly returning to pre-COVID-19 levels, they are still not following typical prepandemic trajectories. ILI is used internationally as a surveillance measure to monitor influenza patterns. In the 4 years postpandemic, and the return of influenza to the NZ environment, ILI continues to be a relatively lower contribution to the overall respiratory burden than prepandemic. This questions the usefulness of ILI as a sole influenza circulation measure, and we suggest combination measures for primary care monitoring are likely to offer more robust patterns for influenza surveillance.

Other acute respiratory presentations have not returned to pre-COVID-19 seasonal peaks. This includes persisting lower presentation rates for young child wheeze illness, otitis media, sore throat and URTIs. This could potentially represent overall a lower viral burden of illness in the NZ community, or perhaps a significant change in health-seeking behaviour by a community who now has a greater awareness of management of respiratory illness.

Presentation by age group shows marked changes through this period. While young children have the highest rates of presentation for respiratory illness to primary care, with the arrival of SARS-CoV-2, rates of presentation for older children and young and mid-life adults were significantly higher than prior to the pandemic and by 2023 mostly returned to the prepandemic pattern. This is likely to be a reflection on the high burden of SARS-CoV-2 for mild and moderately severe disease across the whole population, in contrast to severe disease which particularly affected elderly people.³¹

Strengths

Our use of NLP’s ability to analyse unstructured text in medical records enabled assessment of respiratory presentations across a large cohort over a significant period. Several studies have illustrated the benefits of NLP in pandemic observation, including analysis of Dutch GP notes in predicting clinical outcomes for respiratory patients.³² NLP has been used to examine postacute COVID-19 symptoms in patients with asthma³³ and to enable more efficient triaging of respiratory problems during peak COVID-19 surges in the USA.³⁴

In contrast to the relatively short time frames used in these other studies, our algorithm assessed 10.3 million general practice interactions from 2018 to the end of 2023, spanning before, during and after the COVID-19 pandemic across primary care settings in NZ.

Limitations

Information about the consultation is based on doctor decision making, and the algorithm is limited by the quantity and detail of the recorded information in each consultation in the EHR. Recently, there has been an increasing focus on respiratory illness, including the 2009 H1N1 pandemic, COVID-19 and its aftermath and focus on other virus infections such as RSV. This has provided an impetus for increased recording and detail in the notes, leading to improved accuracy of the NLP algorithm.

The exclusion of out-of-hours consultations may result in an underestimation of primary care respiratory tract-related presentation rates.

NZ does not have the equivalent of a primary care research database, and our results are drawn from ‘routine’ general practice consultations across a wide range of general practice settings. We feel that given the rigour with which the text classifiers for the NLP algorithm have been developed, our results provide results with face validity equivalent to or greater than other commentaries in this area.

Implications for clinical practice and future research

Our study has a focus on the presentation to general practice for coded categories of primary care presentations, rather than the presumed incidence of specific conditions.

This is important, given current diagnostic uncertainty and coding of many presentations. It also aids recognition of the work required when someone presents with symptoms that turn out not to have the positive diagnosis but are still related to it, for example, a COVID-19 negative presentation is still considered COVID-19 workload.

Further algorithm development with this focus on presentation could assist in policy and clinical service planning, particularly exploring trends over time to reveal long-term shifts in patient presentation patterns.

The recent emergence of AI assistive tools for recording and summarising clinical consults is potentially the largest change in clinical notes recording since the introduction of EHRs in primary care and the transition from paper notes. While the use of ‘AI scribes’ had not been introduced during the period of this research, any future studies will need to take into account the features of these AI assistive tools.

Conclusions

The results of this study extend the findings of previous international studies, especially when taking into account NZ’s unique policies and epidemiological context. Many countries saw and have recorded reduced respiratory presentations; this study demonstrates the persistent disruption to seasonal respiratory presentations. The removal of viral circulation does not quickly normalise a return to seasonality.

There has been marked disruption to typical presentation patterns for respiratory illness and continues to be persistent variation from the typical pre-COVID-19 seasonal presentation patterns in primary care utilisation. This variation is driven by a combination of factors, including changes in illness and disease prevalence, as well as shifts in patient and provider behaviour following the pandemic.

This study presents a clear picture of the significant burden of respiratory illness at the primary care level and changing patterns over time. Through the COVID-19 pandemic, the ongoing workload for primary care respiratory work changed in patterns and seasonality but remained very high overall, even though periods of social lockdown and use of Non Pharmacological Interventions (NPIs) reduced community respiratory viral circulation. The results also show that attempts at ‘elimination’ of pandemics do not reduce total primary care workload; they redistribute it.

Our findings highlight that changing pandemic policy or removal of particular policies will still result in multiyear disruption to primary care presentations. It is thus also important to consider that existing methods of ILI surveillance may not perform as well as expected in post-viral elimination contexts. This has important implications for both the management of primary care workloads and public health policy. Understanding baseline utilisation rates may be useful in predicting seasonal variations in primary care visits. Understanding of the prolonged variation in primary care presentation and utilisation patterns can help inform policy to optimise healthcare resource allocation.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.