Understanding the Clinical Implications of Individual Patient Characteristics and Treatment Choice on the Risk of Exacerbation in Asthma Patients with Moderate–Severe Symptoms

Currently, clinical management of patients with asthma is based on a stepwise approach that relies primarily on symptom control; that is, it involves iterations in which symptoms, risk factors, treatment and response are evaluated. Our analysis shows that long-term response to ICS/LABA, which was assessed in terms of future exacerbation risk, is greatly influenced by patients’ characteristics as well as by the treatment choices. It provides insight into the magnitude of the effect of baseline ACQ-5, BMI, FEV₁, smoking status as well as sex and seasonal variation on exacerbation risk. It also allows separation of the treatment effect from these factors. For instance, the cumulative incidence of exacerbations at 1 year was lower for FP/SAL than regular dosing BUD/FOR (8–13%, p < 0.01), where patients with poor control at baseline (ACQ-5 > 1.5) appear to experience the largest effect (Fig. 3). Such a distinction between treatment arms may not be evident in small clinical trials where the sample size affects the precision of the estimated treatment effect and patient baseline symptom control varies across a wide range of ACQ-5 values. Moreover, the use of simulation scenarios indicates that transition to combination therapy with FP/SAL offers a significantly greater reduction in exacerbation risk than higher doses of FP monotherapy. This may be explained by the fact that at therapeutic doses, nearly maximal anti-inflammatory effect is achieved, with further increases in the exposure to ICS having a minor contribution to risk reduction. This is particularly important in those who are overweight or obese, as shown by the difference in the cumulative incidence.

Given that loss of lung function over time in patients with asthma is partially driven by exacerbations, the evidence that differences in BMI, sex and smoking status affect future risk sheds some light onto the ongoing debate about the heterogeneity in individual response to treatment and underlying clinical phenotypes [i.e. type 2 (T2) and non-T2 asthma] [26, 27]. Whilst clinical and demographic factors may not be directly associated with the complex interplay between airway inflammation and airway remodelling, our results suggest that the baseline characteristics identified as covariates on the base hazard parameter describing the risk of exacerbations in moderate to severe asthma could serve as a proxy for airway hyperresponsiveness status.

Even though the benefits of ICS/LABA in this group of patients have been evaluated extensively in different investigations [28‐30], using simulation scenarios it was possible to show that FP/SAL does not only offer acute symptomatic relief but may also have disease-modifying properties, reducing the incidence of exacerbations. Whilst the definition of disease modification may vary, a disease-modifying treatment can be described as a sustained reduction in symptoms and disease activity beyond the immediate or temporary effects of an intervention [31]. Our results showed that initiation of treatment with FP/SAL in patients with poor symptom control not responding to FP monotherapy resulted in approximately 3.2% reduction in the incidence of exacerbations per year relative to those responding to ICS monotherapy. Such an effect appeared to be drug-specific, as the transition to BUD/FOR did not translate into a reduction in exacerbation risk relative to FP monotherapy. This was corroborated by the effect of FP/SAL on symptom control level. Treatment with FP/SAL led to a larger proportion of patients (42%) achieving symptom improvement (i.e. ACQ-5 > 0.5) compared to BUD/FOR (30%).

Finally, our analysis illustrates that baseline characteristics affect base hazard rate and as such contribute to the instantaneous risk but are not predictive of the overall long-term response to an intervention, which is greatly determined by treatment. Indeed, demographic and clinical baseline characteristics in non-responders to FP do not differ significantly from patients on combination therapy. This implies that individual future risk at the time of diagnosis will be miscalculated if only baseline characteristics are used to predict treatment response.

From a methodological perspective, we recognise that to address the main research question from this investigation one should distinguish the contribution of multiple interacting factors to the instantaneous risk of exacerbation, including baseline covariates, trial design and treatment type. Whilst control or stratification of all these factors may not be feasible in a prospective study, CTS offer an opportunity to stratify and eventually offset the effect of confounding or uncontrolled factors [31‐34]. Similarly, any attempt to use retrospective data will be fraught with difficulties, as one needs to consider the effect of censoring and other deviations, which cannot be easily accounted for during data analysis.

Our work has some limitations. Consequently, assumptions had to be made regarding the generalisability of the findings from the different simulation scenarios, as similar protocol conditions may not be easily implemented or controlled in a real-life setting. An overview of the main assumptions and limitations is summarised in Table S4. As a result of ethical reasons, data on the incidence of exacerbation following withdrawal of treatment or use of placebo administration over 12 months is not available. Therefore, a reference intervention had to be defined in order to model the effect of the baseline covariates and treatment choices on exacerbation risk. Moreover, as ICS dose was not a covariate in the TTE model; treatment comparisons were performed using the mean and/or mode dose level used during the maintenance phase of treatment. This was based on the underlying dose–response relationships of FP and BUD [21, 35‐37]. As currently used ICS doses yield nearly maximum pharmacological effect, the impact of varying dose level on basal hazard was assumed to be minor. In fact, here we applied the principles endorsed by Beasley and colleagues [38]. From a statistical perspective, we have also assumed no carryover effect when treatment was switched. In addition, as transition from FP to FP/SAL or BUD/FOR was implemented by design, i.e. switching at a pre-specified time, adjustment or correction in estimates was not deemed necessary [34].

The predicted differences in exacerbation risk reduction following treatment with ICS/LABA combination therapies, which seemed to differ from mean estimates of treatment effect in previously published reports [39‐41], were also carefully scrutinised. A few points need to be considered to understand such a discrepancy. First, it should be noted that randomisation of patients to different treatment arms does not necessarily provide a balanced distribution of all relevant factors that may affect treatment response (i.e. exacerbation events). In fact, baseline risk is not a criterion for inclusion into a trial. Consequently, it is plausible to have cohorts whose baseline characteristics look comparable but have a different exacerbation risk level at baseline. This is illustrated by the heat maps in Fig. 2, which show that the effect of treatment on risk reduction (i.e. reduction in hazard) is comparable across patients with different baseline characteristics. Second, the short duration and relatively small number of events (i.e. exacerbations) observed in each single study may result in imprecise estimates of the treatment effect. These results are often extrapolated to describe annualised exacerbation rates, which may lead to poor generalisability of the estimates to a larger population. By contrast, the model used for the simulations is based on combined trials that have longer duration, allowing prediction of exacerbation rates at shorter or longer intervals, which are likely to be more precise than the results observed in single, shorter clinical trials. Third, the eventual confounding or selection bias associated with the large, but yet potentially limited pool of studies available for the investigation. This has been substantially dismissed following the evaluation of the predictive performance of the model including additional data from 697 patients receiving BUD/FOR (4.5/160 μg two inhalations bid) and 693 patients receiving FP/SAL (50/250 μg one inhalation bid) who were enrolled into study SAM40040 [42]. This study was not available at the time of model development. This additional data increased the sample size of both treatment groups and corroborated the findings in the original analysis, i.e. that the effect of FP/SAL on exacerbation risk is significantly different from BUD/FOR (p < 0.01). The visual predictive check, including the observed exacerbations over 12 months along with the model-predicted 95% confidence intervals, is shown in Fig. S6. This was complemented by a propensity score matching (PSM) [43, 44]. Propensity scores are often used in observational research to construct an artificial control group, which matches subjects with similar propensity scores in both treatment and control, so that potential confounding is reduced. As it can be seen in Fig. S7, comparison of the survival (observed exacerbation events) in a subset of perfectly matched patients reveals similar or slightly larger differences between treatment arms (FP/SAL vs. BUD/FOR). These results show that the findings obtained with the full dataset are unlikely to be caused by confounding. Moreover, this was corroborated by the calculated E-value, i.e. an alternative approach to sensitivity analyses for unmeasured confounding in observational studies [45], which indicates how strong the unmeasured confounding should be to refute the observed results.