General practice registrar evaluation of long COVID in patients presenting with fatigue

Setting

The study was conducted in training practices of the Australian General Practice Education and Training Program.

ReCEnT

This study was nested within the Registrar Clinical Encounters in Training (ReCEnT) project (Magin et al. 2015), an ongoing (2010–present) inception cohort study enrolling GP registrars (vocational trainees in general practice). During the data collection period for this study, ReCEnT was conducted in three of the nine then-extant Australian general practice training regions covering the Hunter, Manning, Central Coast, and New England regions of New South Wales (Region 1), Tasmania (Region 2), Eastern Victoria (Region 3), Sydney (Region 4), and the remainder of New South Wales and the Australian Capital Territory (Region 5). Collectively this included 44% of Australian GP registrars.

ReCEnT collects data on registrars’ in-consultation clinical and educational experiences and actions. Data are recorded via a dedicated online portal for 60 consecutive consultations during each of a registrar’s three 6-month mandatory community-based training terms. Included consultations are clinic-based face-to-face and telehealth encounters, each recorded on an individual case report form (CRF). Home visits and residential aged care facility visits are excluded. In-consultation data are analysed along with registrar- and practice-level data collected via questionnaires that registrars complete prior to commencing in-consultation data collection.

Data for the current study were collected during training semesters in the second half of 2022 and first half of 2023.

Outcome variables

The main analysis outcomes were two dichotomous variables derived from responses to a CRF item that registrars were asked to complete if fatigue was a presenting symptom or elicited during the consultation. This item asked registrars to rate how likely it was that long COVID was the cause for the patient’s fatigue. The response options comprised a Likert scale (‘Very unlikely’, ‘Unlikely’, ‘Likely’, ‘Very likely’) and the option to indicate that they did not consider long COVID as a differential (‘Didn’t consider’). Two dichotomised outcomes were used in analyses:

Predictor variables

Predictor variables used in the multivariable analyses were at the patient, registrar, practice, and consultation level.

Practice-level variables were size (large practice >5 full-time-equivalent GPs); whether the practice was fully bulk-billing (accepts the government rebate as full payment for all consultations); training region; rurality/urbanicity (via the Australian Standard Geographical Classification – Remoteness Area (ASGS-RA)) (Australian Government Department of Health and Aged Care 2021); and socio-economic status of the practice location (via the Socioeconomic Index for Area Relative Index of Disadvantage (SEIFA-IRSAD)) (Australian Bureau of Statistics 2016). ASGS-RA and SEIFA-IRSAD were defined by practice postcode.

Patient-level variables were age; gender; Aboriginal or Torres Strait Islander status; being of a non-English speaking background (NESB); and if the patient was new to the practice and/or the registrar.

Registrar-level variables were age; gender; term of training; part-time/full-time training status; worked at the practice in a previous term; and location of first medical degree (Australian medical graduate (AMG) or international medical graduate (IMG)).

Consultation-level variables were the duration of consultation (in minutes); the number of problems addressed; and whether the consultation was via telehealth.

Statistical methods

Analysis was at the level of individual consultation.

Descriptive statistics were calculated as proportions with 95% confidence intervals (CIs) for all non-missing observations of a given variable. These included the proportion of all consultations where fatigue was presented or elicited as a symptom; the proportion of these fatigue consultations in which long COVID was considered; and the proportion of these ‘long COVID considered’ consultations in which long COVID was thought the likely diagnosis. For all predictor variables, frequencies (categorical variables) or mean and standard deviation (s.d.) (continuous variables) were calculated.

For the main analyses, univariable and multivariable logistic regression was used with the generalised estimating equations framework to account for repeated observations within registrars. An exchangeable working correlation structure was assumed and missing values were handled with complete case analysis. For the regressions with outcome variable ‘long COVID considered’, the analysis was restricted to encounters where fatigue was considered a symptom. For the regressions with outcome ‘long COVID thought likely’, the analysis was restricted to encounters where fatigue was a symptom and long COVID was considered in the differential diagnosis. An augmented backwards selection process was followed. An initial working set of variables was defined as the 20 potential explanatory variables and a logistic model fitted to this working set. Variables with P-values > 0.2 were tested for removal. A variable was removed if the resulting model did not have substantively different effect sizes than the previous model. A substantively different effect size was defined as being more than about 10% difference in β coefficients for other variables in the model. The process was repeated until no further variables were identified for removal. Model fit was assessed using the Hosmer–Lemeshow (H–L) goodness of fit test applied to a standard logistic model. The assumption of linearity in the log-odds for continuous variables was also checked. Predictive accuracy was assessed with the C-statistic.

The regressions modelled the log-odds for ‘long COVID considered’ and ‘long COVID thought likely’. Effects are expressed as odds ratios (ORs) with 95% CI. Significance was declared at the conventional 0.05 level, with the magnitude and precision of effect estimates also used to interpret results. Analyses were performed using STATA ver. 18.0 (StataCorp, College Station, Texas, United States) and SAS ver. 9.4 (SAS Institute Inc., Cary, North Carolina, United States).

Ethics approval

Ethics approval was granted by the University of Newcastle Human Research Ethics Committee (reference H-2009-0323).

Outcome ‘long COVID considered’

Of the 3193 encounters in the study period in which fatigue was a presenting or elicited symptom, the registrar considered long COVID as a differential in 2563 encounters (80.27%, 95% CI 78.85–81.64%).

The characteristics associated with ‘long COVID considered’ are presented in Table 2. The univariable and multivariable logistic regression models for the outcome ‘long COVID considered’ are presented in Table 3.

On multivariable analysis, older registrars (OR 1.05 per year [95% CI 1.02–1.08], P = 0.001) and the consultation being via telehealth (OR 1.74 [95% CI 1.30–2.31], P < 0.001) were associated with considering long COVID as a differential diagnosis. There was some evidence (P = 0.10) for the patient being from a non-English speaking background being associated with it being less likely for long COVID to be considered (OR 0.73 [95% CI 0.51–1.06]).

No variables were removed during model selection. The non-significant H-L P-value of 0.17 suggested satisfactory model fit, and the C-statistic of 0.64 indicated fair predicative accuracy. There did not appear to be any violation of linearity in the log-odds for the continuous variables.

Outcome ‘long COVID thought likely’

Of the 2563 encounters in which long COVID was considered, the registrar considered long COVID to be the likely cause for the patient’s fatigue in 465 encounters (18.14%, 95% CI 16.67–19.69%).

The characteristics associated with ‘long COVID thought likely’ are presented in Table 4. The univariable and multivariable logistic regression models for the outcome ‘long COVID thought likely’ are presented in Table 5.

On multivariable analysis, the practice being routinely bulk-billing was associated with COVID being thought likely (OR 2.60 [95% CI 1.61–4.18], P < 0.001), as was large practice size (OR 0.70 for small practice size [95% CI 0.51–0.97], P = 0.033). The consultation being conducted by telehealth (OR 2.33 [95% CI 1.65–3.30], P < 0.001) was associated with COVID being thought likely, as was the registrar being an IMG (OR 0.58 for Australian graduates [95% CI 0.39–0.88], P = 0.010). There were also significant associations with region and (with clinically small effect size) shorter consultation duration.

Model selection removed the consultation variable ‘number of problems’. The non-significant H–L P-value of 0.16 suggested satisfactory model fit, and the C-statistic of 0.69 indicated fair predicative accuracy. There did not appear to be any violation of linearity in the log-odds for the continuous variables.

Summary of main findings

This study aimed to examine whether GP registrars have adapted their diagnostic approach to fatigue to include long COVID. We found that registrars considered long COVID as a differential in 80% of presentations involving fatigue, indicating that registrars have indeed adapted their approach. Registrars thought long COVID was a likely diagnosis in 18% of fatigue encounters, suggesting that they implicitly considered long COVID to be about as prevalent among all patients presenting with fatigue as it is among all patients who have had recent acute COVID, during a time at which COVID restrictions in Australia had largely been lifted and community transmission was widespread (Australian Government Department of Health and Aged Care 2024).

No patient variables were significantly associated with either outcome in the two logistic regression models. While this suggests that patient factors may not significantly influence registrars in considering long COVID, it also indicates that patient demographic factors well-established to be associated with the risk of long COVID, including age and sex (Tsampasian et al. 2023), were not reflected in registrars’ decision-making. The inclusion of only patients presenting to their GP with fatigue rather than all patients with a previous acute COVID-19 infection or with potential long COVID symptoms may mean that the relationships between patient factors and long COVID incidence within this sample differed from that in the broader population. Several proposed long COVID clustering or subtyping schemes include fatigue as a key differentiating symptom (Yong and Liu 2022; Reese et al. 2023), and there is evidence that patient sex relates to symptomatology including fatigue (Sylvester et al. 2022). Alternatively, this may reflect small true effect sizes for these patient demographic factors on long COVID incidence, which this design lacked power to capture.

Telehealth was notable as a factor significantly and strongly positively associated with both model outcomes, with an adjusted OR of 1.74 (95% CI 1.30–2.31) for ‘long COVID considered’ and an adjusted OR of 2.33 (95% CI 1.65–3.30) for ‘long COVID thought likely’. Driven by the need to provide care in the context of lockdown restrictions or to patients with potential acute COVID-19 symptoms, the rapid expansion of telehealth consultations in Australian general practice began early in the COVID-19 pandemic, was supported by permanent changes to the Medicare Benefits Schedule (MBS), and has since become a permanent part of practice (Fisher et al. 2022). The strong association with telehealth may simply reflect that patients with long COVID or symptoms such as persistent cough or dyspnoea are more likely to seek telehealth consults, due to practice rules around patients with respiratory symptoms or patients’ personal choice. Patients have identified reduced travel time, convenience, and rapid access as some of the perceived benefits of telehealth (Ward et al. 2022), and these benefits may be more appealing to patients suffering with long COVID-related post-exertional malaise (Davis et al. 2023). However, the strong association between telehealth and ‘long COVID thought likely’ raises the question of whether a telehealth consult is sufficient to thoroughly exclude alternative explanations for a patients’ symptoms, a diagnostic criterion under the World Health Organization case definition (World Health Organization 2021). Registrars may face increased diagnostic uncertainty when they cannot see the patient or perform a physical examination, which may reduce their confidence in alternative diagnoses. Further research into how GPs are approaching potential long COVID presentations and their diagnostic process could clarify the role of telehealth in this condition.

Two registrar factors were significantly associated with the model outcomes. There was a significant positive relationship (OR 1.05 [95% CI 1.02–1.08]) between registrar age and ‘long COVID considered’, perhaps reflecting a broader diagnostic approach with increasing clinical experience and knowledge. There was a significant and strongly negative relationship (OR 0.58 [95% CI 0.39–0.88]) between a registrar having qualified as a doctor in Australia and ‘long COVID thought likely’. This cannot be attributed to differences in medical education on COVID-19 and long COVID, as the timeline for GP training in Australia means all registrars participating in ReCEnT 2022 and 2023 would have completed their medical education prior to the pandemic, with the average registrar graduating in 2015 (Table 1). It may represent broader differences in training and experience, or systematic differences in clinical setting, patient population, and other factors driven by regulatory restrictions on where non-Australian qualified registrars can practice (though many of these differences were adjusted for in the multivariable analyses).

Three practice factors were significantly associated with ‘long COVID thought likely’, while none were significantly associated with ‘long COVID considered’. Practice routinely bulk billing was associated with a substantially higher odds (OR 2.60 [95% CI 1.61–4.18]) of ‘long COVID thought likely’. Bulk-billing is associated with a number of geographic, practice, and patient factors not captured in this study, including whether a patient attends multiple GP practices (Glenister et al. 2021). Smaller practices and practices in large capital cities were also associated with a significant reduction in odds for ‘long COVID thought likely’. These associations may represent differences in underlying long COVID incidence in the populations served by these practices, for example due to differences in factors associated with long COVID risk (e.g. comorbidities, smoking status, vaccination status) that were not captured in this study. Differences in practice driven by resources and access to services, for example to referral pathways for long COVID, may also have indirectly influenced registrars’ diagnostic process.

Strengths and limitations

This study benefits from a large and geographically diverse dataset of clinical encounters in Australia. The large number of potential confounding variables measured in the study allows for fine-grained adjustment for confounding. The response rate is very high for a study of GPs.

A limitation of this study is that it includes data only from GP registrars rather than all practising GPs. Future work would benefit from a broader sample. The iterated design risks bias due to priming and the Hawthorne effect, with registrars perhaps more likely to consider long COVID having answered the post-encounter questions about long COVID for a previous fatigue presentation, or because they become aware this is being measured. This is mitigated by the presence of many other questions on the CRF (Davey et al. 2022). The study was also limited in collecting data only from presentations with fatigue, rather than all potential long COVID symptoms, and not collecting data on whether and when a patient had an acute COVID-19 infection.

Future research

Future research examining more experienced GPs and symptoms other than fatigue, including qualitative inquiry into GPs’ diagnostic processes and in particular examining face-to-face compared to telehealth consultations, would be of benefit to more completely understand GP approaches to long COVID. It would also be of benefit to link or otherwise capture data on whether and when a given patient had an acute COVID-19 infection, as this would contextualise the GP’s assessment of the risk of long COVID.