Dynamics of inhaled corticosteroid use are associated with asthma attacks

Inhaled corticosteroids (ICS) suppress eosinophilic airway inflammation in asthma, but patients may not adhere to prescribed use. Mean adherence—averaging total doses taken over prescribed—fails to capture many aspects of adherence. Patients with difficult-to-treat asthma underwent electronic monitoring of ICS, with data collected over 50 days. These were used to calculate entropy (H) a measure of irregular inhaler use over this period, defined in terms of transitional probabilities between different levels of adherence, further partitioned into increasing (Hinc) or decreasing (Hdec) adherence. Mean adherence, time between actuations (Gapmax), and cumulative time- and dose-based variability (area-under-the-curve) were measured. Associations between adherence metrics and 6-month asthma status and attacks were assessed. Only H and Hdec were associated with poor baseline status and 6-month outcomes: H and Hdec correlated negatively with baseline quality of life (H:Spearman rS = − 0·330, p = 0·019, Hdec:rS = − 0·385, p = 0·006) and symptom control (H:rS = − 0·288, p = 0·041, Hdec: rS = − 0·351, p = 0·012). H was associated with subsequent asthma attacks requiring hospitalisation (Wilcoxon Z-statistic = − 2.34, p = 0·019), and Hdec with subsequent asthma attacks of other severities. Significant associations were maintained in multivariable analyses, except when adjusted for blood eosinophils. Entropy analysis may provide insight into adherence behavior, and guide assessment and improvement of adherence in uncontrolled asthma.

. Asthma that remains uncontrolled despite the use of high-dose ICS-based controller is regarded as 'difficult-to-treat' , and presents a complex clinical challenge 3 . Medication adherence describes the extent to which patients use medication as prescribed 4,5 . In difficultto-treat asthma, patients frequently deviate from prescribed use 6,7 . This can be tracked objectively by attaching electronic monitoring devices (EMDs) to a patient's inhaler, recording the date and time of each actuation 8 . In future, electronic monitors may be routinely integrated into inhalers during manufacture 9 .
Clarifying the interplay between poor adherence and adverse outcomes could help improve adherence and enhance patient health. So far, work has focused on time-averaged metrics, typically mean adherence (total doses taken/total doses prescribed) and it has been difficult to demonstrate a relationship between asthma outcomes and mean adherence 10,11 . This averaged metric fails to capture potentially important variations in medicationtaking behaviour, e.g. a mean adherence rate of 50% cannot distinguish between one patient consistently taking half the prescribed dose daily and another taking the full prescribed dose, but for only half the required period. Other metrics do take into account either the interval between doses, or the time above a minimum dose threshold, and some have shown a relationship to attack rates in airways diseases 10,12 . However, these strategies still only represent summative time-averaged metrics, and do not describe day-to-day deviations from regular prescribed usage.
We designed adherence metrics to capture via EMD the extent to which patients with difficult-to-treat asthma deviate from regular controller usage, by measuring the entropy-irregularity, or disorder-with which daily medication doses are taken. The concept of entropy is derived from information theory where it is used to quantify the 'information' in a process. Entropy has been previously applied to respiratory symptoms 13 , breathing patterns, and lung function 14 . We examined whether these entropy measures of adherence related to specific patient characteristics or predicted subsequent asthma-related clinical outcomes. For comparison, we also measured conventional mean adherence, time-and dose-based variability (using additional metrics reflecting missed days and incomplete doses respectively), and the duration of gaps in which patients completely forwent medication.
We hypothesised that the degree of irregularity of ICS controller usage may be more relevant in difficultto-treat asthma, and better predict poor outcomes. Highly disordered medication-taking behaviour may place patients at higher clinical risk and may be associated with poorer outcomes.

Methods
Study participants. Our tertiary centre receives referrals of adults with difficult-to-treat asthma from specialists in secondary care 17 . Patients underwent multidisciplinary assessment according to a pre-specified protocol over three visits over six months-previously reported in detail [18][19][20][21][22] -to confirm asthma diagnosis, address comorbidities, and optimise treatment (Fig. 1).
Inhaler technique was reviewed and optimised.
Electronic devices. A compatible EMD (Adherium, Auckland, New Zealand) was fitted to the patient's ICS-containing controller inhaler at visit one (day 0) and EMD-collected ICS data were uploaded at visit two Adherence metrics. Metrics were quantified using Python (Python Software Foundation, version 3·6). For standardisation, the first 50 days of available data were extracted for each patient (Fig. 1); this excluded any days with missing data (defined as days when inhaler was not attached-logged by the device as distinct to zero adherence). As the EMD was returned at visit two, adherence data were not available to day 180. Last observation carried forward was not performed to minimise the risk of introducing bias into the adherence metrics, particularly entropy. Entropy (H), a measure of disorder, was adapted to the adherence data to reflect the various ways in which the patients changed their ICS-taking behavior from day to day. In information theory, a 'Markov chain' can be used to describe the sequence of occurrences of certain 'states' and the probabilities of transitioning from one state to another given the previous state (e.g. the appearance of specific sequences of letters in a message). H is then used to quantify the complexity of the information, in terms of the transitional probabilities between states, for all possible states observed. Here, we classified adherence into different levels, which are analogous to the different states of a Markov chain, similar to an approach previously applied to respiratory symptoms 13 . Given an adherence time series x, where x is the dose taken/prescribed dose expressed as a percentage, we obtained a state-based series by mapping each element of x i to the state space S = {1,2,3,4,5,6,7} as follows:

State, si
Dose range We then computed the 7 × 7 transitional probability matrix, which comprised probabilities P i,j denoting the probability of transitioning from state j, given an initial state i, for every combination of states i,j. The entropy (H) of the system was determined as H = N−1 i,j=0 P i,j (−ln(P i,j )) , representing the disorder of transitions between daily dose states. An example of H calculation is shown in Fig. 2.
We further partitioned the probability matrix into "increasing" (H inc ) or "decreasing" (H dec ) adherence, by only considering transitions that moved from lower to higher adherence states on the next available day, or vice versa, respectively, i.e. splitting the transitional probabilities along the diagonal of the matrix. H inc was determined as N−1 i,j=0 P i,j (−ln(P i,j )) , where i < j. Similarly, H dec was defined as N−1 i,j=0 P i,j (−ln(P i,j )) , where i > j. Thus, while H represents the day to day changes in adherence levels in general, H inc represents the different ways in which a patient may increase their adherence level, and conversely H dec the different ways in which decreases in adherence may occur. Figure 3 illustrates how different adherence time series with the same conventional mean adherence (PT mean , see below) may vary in H inc and H dec .
Conventional mean adherence was described as PT mean , expressed as a percentage of the prescribed number of puffs per day. This was capped at 100%-all higher daily instances were converted to 100% to allow comparison of the mean (PT mean,cap ) with other published studies. PT SD and PT CV represented the standard deviation and coefficient of variation of PT mean respectively, based on uncapped data to capture the full variability in adherence.
Area under the curve (AUC) measures were inspired by methods previously described 10 to investigate both time-and dose-based adherence variability. In brief, perfect time adherence was first defined as medication taken daily (over the first 50 days), regardless of dosage. The time adherence curve was then defined as the cumulative www.nature.com/scientificreports/ sum of every day when medication was taken. Thus, perfect time adherence corresponded to a straight line with an area under the curve normalised to 100%. The time-based AUC (T-AUC) for an individual patient was taken as the difference between their time adherence curve and the perfect curve, expressed as a percentage deviation from 100%. In this way, the T-AUC described how consistently the patient took any medication over time. The use of the cumulative sum meant that earlier and/or larger gaps have greater effects on the T-AUC. Similarly, the dose-based area AUC (D-AUC) described the cumulative deviation from the patient's total prescribed dose over the same 50-day period. We also multiplied the time-based deviation and the dose-based adherence for each day, to construct a composite curve. The Prod-AUC was then calculated as the cumulative deviation from the product of the perfect time x dose curves. This metric thus reflects adherence behavior in terms of both time and doses taken over the given time period. The Gap max metric described the maximum length of gaps between days when medication was last taken, regardless of number of puffs within a day, during the 50-day period.

Statistical analysis. Relationships between adherence metrics and clinical characteristics (at baseline) and
asthma outcomes (at/over six months) were examined using Spearman rank correlation (r s ) for continuous variables, and Wilcoxon rank sum or Kruskal-Wallis tests for comparisons between 2 groups or > 2 groups, respectively. Multivariable regression was performed to confirm if any significant associations between adherence metrics and clinical characteristics and asthma outcomes identified from univariate analyses were still independent predictors after adjusting for the potential confounders of age, sex, baseline eosinophils and baseline lung function. Adjustment for baseline asthma questionnaire scores (AQLQ and ACT) was undertaken where the respective asthma questionnaire score was the outcome.

Results
Participants. Systematic assessment was undertaken by 108 patients. Forty (37%) did not receive a monitoring device: four (3.7%) from physician choice, two (1.9%) declined, two (1.9%) did not have asthma (and were diagnosed with vocal cord dysfunction), and 32 (29%) had inhalers with no compatible EMDs available in Australia. Among 68 patients who underwent inhaler monitoring, 11 (16.2%) devices had less than 50 days of data due to device detachment, and 4 (5.9%) devices malfunctioned (Fig. 4). These patients were excluded from the analysis. Of the 11 devices with missing data, the mean number of days monitored was 37, (SD 69, range 3-220 days). Data were analysed from 53 patients (Table 1).

Adherence metrics and baseline clinical characteristics. Summary statistics of adherence metrics
calculated for the first 50 days in all patients are reported in supplemental Table S1. No adherence metric was related to age, sex or baseline lung function on correlation testing (Supplemental Table S2). PT mean,cap correlated with baseline asthma quality of life as measured by AQLQ (Spearman correlation, r S = 0·284, p = 0·046). Large gaps in inhaler use (Gap max ) and lower T-AUC were associated with a greater likelihood of previous intensive care or hospitalisation for an asthma attack in the six-month period prior to visit one (Wilcoxon rank sum test), Gap max : Z = − 2.068, p = 0·039 and Z = − 2.08, p = 0·037 respectively, T-AUC: Z = − 2.065, p = 0·039 and Z = − 2.042, p = 0·041 respectively). Regarding entropy, higher H correlated negatively with baseline AQLQ and ACT scores (r S = − 0·330, p = 0·019 and r S = − 0·288, p = 0·041 respectively). Higher H dec similarly correlated negatively with baseline AQLQ and ACT (r S = -0·385, p = 0·006 and r S = − 0·351, p = 0·012 respectively), and was further associated with higher SABA reliever use in terms of puffs and days per week (r S = 0·318, p = 0·02 and r S = 0·286, p = 0·04 respectively).
The relationships between entropy measures (H and H dec ) and baseline ACT and reliever use remained significant following multivariable regression models adjusting for age, sex, baseline eosinophil count and baseline FEV, while all other measures did not (Table 2).
Adherence metrics and subsequent outcomes at six months. Among all adherence metrics, only measures of entropy, measured in the first 50 days, were associated with asthma outcomes at six months (Supplemental Table S3). Higher H was associated with more asthma attacks requiring hospitalisation over six months  Fig. 5a).Higher H dec was associated with more asthma attacks over the six months prior to visit three, requiring a visit to a general practitioner (Z = − 2.43, p = 0·015), oral corticosteroids (Z = − 2.508, p = 0·012), or hospitalisation (Z = − 2.07, p = 0·038, Fig. 5b-d). (All comparisons performed by Wilcoxon Rank Sums Test). For regression analysis, it was only possible to adjust for one confounder at a time, due to the amount of patient data available at 6 months. Relationships between H and asthma attacks requiring hospitalisation remained significant regardless of adjustment for age, sex and baseline FEV 1 (Table 3). Similarly, relationships between H dec and attacks requiring oral corticosteroids or general practitioner visits also remained significant with these adjustments (Table 3) and also when adjusted for baseline number of attacks in the six months prior However, after adjustment for baseline eosinophil count, the relationships between H or H dec and attacks at six months were no longer significant. Further examination showed that H (but not H dec ) was correlated with baseline eosinophil count (r S = 0·352, p = 0·045) suggesting collinearity between baseline eosinophils and H.

Discussion
With increasing emphasis on inhaler adherence monitoring in airways diseases, particularly in the era of biologic therapies for severe asthma, there is a pressing need to identify the optimal metrics with which to measure inhaled controller adherence 7,26,27 . We showed that disordered controller use in difficult-to-treat asthma-as reflected by entropy analysis-reflected poor baseline asthma control and were associated with subsequent attacks of any severity. This could potentially be mediated through unchecked eosinophilic inflammation. Entropy measures have previously been used to describe respiratory symptoms and breathing patterns, with higher entropy associated with adverse outcomes 13,28 . We designed entropy measures (H, H dec and H inc ) to measure the irregularity of day-to-day dose-taking behaviour, analogous to the original use of H in information theory to quantify the complexity in strings of text 29 . We used it to describe the diversity in patterns in observed transitions in adherence, choosing to also examine irregularity or diversity in increases and decreases in adherence, as they may be clinically relevant. In considering all (or a subset in the case of H dec and H inc ) of the elements in the transitional probability matrix, our method of calculating H differs from that of Usemann et al., where entropy was calculated from rows of elements and then averaged 13 . Nevertheless, H calculated using our method was highly correlated with their method when applied to this dataset (r = 0·912, p < 0·001, data available on request).
To accommodate the original study design, we chose a period of 50 days to maximise participants with sufficient data. This proof-of-concept study justifies validation in larger cohorts and the development of more dynamic measures of entropy, similar to our previous work on peak flows to predict attacks 30 .
Entropy metrics, specifically in relation to decreasing states (H dec ) over a 50-day period, related to worse asthma control and increased short-acting reliever use at baseline. Notably, greater H dec also predicted subsequent risk of attacks of any severity, whether requiring general practitioner visit, increase in oral steroids, or hospitalisation (the latter also predicted by H). That H dec , rather than H inc , has these relationships suggests that irregular drops in adherence may have more clinical impact than over-adherence. These relationships were no longer significant when adjusting for baseline peripheral eosinophil levels. The correlation between H (though not H dec ) and peripheral eosinophils suggests that higher baseline eosinophil counts may represent previous poor adherence. We hypothesise that the same pattern of behaviour may then have continued during the period of monitoring, with unchecked eosinophilic inflammation subsequently leading to asthma attacks. Previous studies have demonstrated that peripheral eosinophils are an independent predictor for asthma attacks 31 . Within this small study, baseline blood eosinophils did not predict asthma attacks at 6 months, nor was FeNO related to any adherence measures (supplemental data).
While non-adherence can be intentional due to issues such as mistrust, lack of medication understanding, fixed beliefs and cost, unintentional disordered medication use may also indicate a corresponding degree of chaos in patients' lives. In asthma, poor family routines accompanied diminished inhaler adherence in children 32 . In post-myocardial infarct patients, 'life-chaos'-a highly variable daily routine with an inability to plan and anticipate the future, paralleled poor adherence to cardiac medication 15 . Similar life-chaos among patients with HIV was associated with increased health care use and missed clinic appointments 16 . We speculate that the extent of entropy in controller use in difficult-to-treat asthma may also reflect overall life-chaos. Measurement of entropy in inhaled controller use could be used in the clinic setting to target patients particularly with high H dec for adherence interventions. Such patients may have otherwise been missed if conventional averaged adherence measures were used (Fig. 3). Entropy measures may also prompt the clinician to review the wider social situation of the patient for other indicators of 'life chaos' .
As anticipated, conventional mean adherence (PT mean,cap ) in our study (following adjustment for potential confounders) was not related to baseline asthma status, nor predicted longitudinal outcomes.
Similarly, neither variability in dosage nor timing metric was associated with clinically important outcomes (Supplemental tables S2 and S3). In a previous analysis of a clinical trial in moderate asthma, the use of AUCbased metrics did relate to asthma-related quality of life and lung function by peak flow measurement 10 . Note that our AUC metrics were based upon, but were not directly comparable to previously-published methods 10 , which accounted for technique/device errors using a specialised INCA device. Furthermore, our study cohort included consecutive patients drawn from clinical practice. Table 3. Multivariable analysis relating adherence metrics to 6 month outcomes. GP general practitioner. Adherence metrics and outcome measures reported here are those which showed significant associations in univariate analyses. www.nature.com/scientificreports/ Our patient population had significant disease with poorly controlled symptoms and high exacerbation rates, despite having previously been assessed by respiratory specialists. We have previously demonstrated that this population still has high non-adherence rates despite specialist intervention. Our results are likely to be representative of difficult-to-treat asthma patients encountered in the 'real world' , but may not represent less severe patients. The association of entropy with other behaviour that can affect adherence such as mistrust of medication, financial barriers, and not attending a pharmacy access to refill prescriptions would be worth pursuing with future research.

Limitations
Given the complexity of difficult-to-treat asthma, poor disease control may relate to a wide range of disease and patient factors, e.g. biological severity, corticosteroid insensitivity, multimorbidity, poor self-management skills-all addressed in our clinic's systematic protocol [33][34][35] . Notwithstanding the presence of such confounding issues, a significant effect of disordered controller use on risk of asthma attack remained detectable. However, it is possible our single-center study had insufficient statistical power from a reduction in data available due to device incompatibility device malfunction, missing data, small sample size and short duration of data collection, to detect weaker associations. We also relied on patient recollection for asthma attack history which could be inaccurate, although these data were verified when available in medical records. We explored a range of metrics, baseline characteristics, and asthma outcomes, so increasing the likelihood of a chance finding. However, the consistent pattern of results and their persistence following adjustment for confounding both support a true result. We only collected adherence data between visit one and two of our study (most consistently for 50 days), and analysed outcomes at day 180 (visit three). It is possible that adherence would have improved beyond 50 days, however we wished to analyse the impact of the patient's initial adherence behaviour on future asthma outcomes. It is likely other aspects of adherence behaviour would add to the predictive power of entropy measures; larger validation datasets would enable further exploration as well as control for other possible confounders in the same model. Future studies could also explore the impact of patient socioeconomic status or device polypharmacy on entropy of inhaled controller use as well as examine aspects of 'life chaos' more qualitatively.

Conclusions
We showed higher irregularity assessed by entropy in controller use of patients with difficult-to-treat asthma, with effects that appear mediated through eosinophilic inflammation, and were associated with an increased risk of future attacks. Entropy may reflect the 'life chaos' experienced by people with difficult-to-treat asthma, a possible target for appropriate intervention. Entropy analysis could guide future approaches to improve adherence and enhance patient health, potentially applicable to other domains of respiratory or other chronic disease.