Longitudinal Patterns and Patient-Reported Barriers to Maintenance Inhaler Adherence in COPD: A Three-Wave Observational Study Using Routine Follow-Up Records

Abstract

Purpose

To describe short-term patterns of maintenance inhaler adherence and patient-reported barriers among adults with chronic obstructive pulmonary disease (COPD), and to examine whether barriers documented at earlier follow-ups were associated with subsequent adherence.

Methods

This single-center longitudinal observational study included 278 adults with COPD followed at approximately 1-month intervals for three waves. Adherence was classified from patient- or caregiver-reported inhaler use obtained through routine semi-structured follow-up during the preceding month; unclear or unassessed use was coded missing. Transparent text rules identified five barrier categories and were validated against independent review of 150 randomly sampled records. The primary temporally ordered analysis used Firth logistic regression to compare Wave 3 low adherence between participants with any barrier documented during Waves 1–2 and those with none. Barrier-category count, adjacent-wave models, contemporaneous generalized estimating equations, and inverse probability-of-observation weighting were secondary or sensitivity analyses.

Results

Adherence was evaluable for 264, 247, and 236 participants at Waves 1, 2, and 3; high adherence was recorded in 259 (98.1%), 222 (89.9%), and 189 (80.1%), respectively. Medication beliefs/perceived effectiveness were the most frequently documented barrier across any wave (35/278, 12.6%). In the primary analysis, any barrier documented during Waves 1–2 was associated with low adherence at Wave 3 compared with no prior barrier (adjusted odds ratio [OR] 4.29, 95% confidence interval [CI] 2.00–9.16). The corresponding result using truncated stabilized observation weights was OR 4.81 (95% CI 2.20–10.50). In a supportive Wave 2→3 model, current high adherence was associated with next-wave high adherence (OR 19.55, 95% CI 3.03–284.84), whereas individual barrier estimates were imprecise.

Conclusion

High adherence was recorded less frequently at later follow-ups, and the presence of a previously documented barrier was associated with low adherence at Wave 3. These observational findings do not establish causation or individual prediction.

Plain Language Summary

People with COPD often need to use maintenance inhalers regularly. We reviewed three routine follow-ups, approximately 1 month apart, for 278 patients. Staff asked patients or their main caregivers how often they had used the prescribed inhalers during the previous month. We also used transparent text rules to identify five types of difficulty recorded in the notes: beliefs about the need for or benefit of treatment, concerns about side effects, problems obtaining or refilling medication, forgetfulness or disrupted routines, and difficulty using the inhaler device.

We recorded high adherence less often at later follow-ups. We found that patients with at least one barrier documented during the first two follow-ups were more likely to have low adherence at the third follow-up. We conducted the study at one hospital, and patients or caregivers reported adherence rather than using electronic monitoring. Some outcomes were missing, and several barriers were uncommon. These limitations mean that the findings should be interpreted as associations rather than evidence of causation or individual prediction. Clinicians should confirm any automatically identified concern through discussion with the patient or caregiver.

Introduction

Long-term inhaled therapy is central to stable COPD management, but potential benefit depends partly on whether patients initiate treatment, use it as intended, and persist over time.^1–3 Reported adherence varies markedly with population, data source, recall period, and threshold.^4–8 Inhaler treatment also combines medication-taking behavior with device operation, so refill measures, questionnaires, self-report, and electronic monitoring capture different aspects of adherence.^9–13

Barriers to maintenance inhaler use include beliefs about necessity or effectiveness, concerns about adverse effects, forgetfulness and competing routines, cost or refill difficulties, and problems operating the device.^4,6–13 Adherence and persistence may change during treatment,^14–16 yet many studies assess barriers at a single time or through a fixed questionnaire. Routine follow-up notes may provide time-specific context about missed use, self-directed dose reduction, discontinuation, and perceived difficulties, but the information is unstructured. Identifying why inhaler use differs from the prescribed regimen may be clinically useful because different problems may require different follow-up conversations or forms of support. For example, medication beliefs, practical access difficulties, disrupted routines, and device-use problems represent distinct issues that may not be addressed in the same way. However, identifying a documented barrier does not by itself establish which intervention will be effective for an individual patient.

Clinical text methods can identify adherence information, although negation, clinician-directed treatment changes, uncertainty, and local expressions can cause misclassification.^17–19 For a modest Chinese-language dataset, transparent rules are auditable and can be checked against source records, but they should be treated as an extraction aid rather than an autonomous clinical classifier.

We used three waves of routine COPD follow-up records to describe reported maintenance inhaler use, adherence-state transitions, and five patient-reported barrier categories. We separately examined same-wave associations, adjacent-wave associations, and whether any barrier was documented before Wave 3, with the number of distinct prior barrier categories examined exploratorily. The study was designed to estimate associations and generate hypotheses, not causal effects or individual-level predictions.

Patients and Methods

Study Design and Participants

This single-center longitudinal observational study used routine follow-up data from patients consecutively entering the COPD management pathway at Haiyan People’s Hospital between Jan 1, 2025 and June 30, 2026. COPD was diagnosed according to Global Initiative for Chronic Obstructive Lung Disease criteria: post-bronchodilator forced expiratory volume in 1 second to forced vital capacity ratio below 0.70 in the appropriate clinical context. Three follow-up waves were scheduled at approximately 1-month intervals. During the study period, 280 patients entered the COPD management pathway. Two patients completed fewer than two follow-up assessments and were excluded, leaving 278 participants in the final analytic cohort (Supplementary Figure S1).

Ascertainment of Maintenance Inhaler use Adherence

At each wave, trained follow-up staff conducted a routine semi-structured follow-up using the same core inquiry domains. The same operational classification rules were applied across all three follow-up waves. The wording and order of questions could vary according to the patient’s or primary caregiver’s responses and the clinical context, but staff consistently sought information on whether the prescribed maintenance inhaler was being used, the reported frequency of use during the preceding approximately 1 month, missed or irregular use, patient-initiated dose reduction or discontinuation, and whether any treatment change had been directed by a clinician. Additional questions addressed the reasons for nonadherent use when reported, including medication beliefs or perceived lack of effectiveness, side-effect concerns, access or refill difficulties, forgetfulness or routine disruption, and inhaler-device difficulties. Information obtained during the follow-up was summarized in a narrative clinical note. The domains covered by the routine follow-up are provided in Supplementary Appendix S1.

Adherence was classified qualitatively by comparing the prescribed regimen with the reported implementation during the preceding approximately 1 month. High adherence indicated reported use consistent with the prescribed regimen without documented recurrent or irregular missed use, patient-initiated dose reduction, or patient-initiated discontinuation. Low adherence indicated documented self-directed deviation from the prescribed regimen, including recurrent or irregular missed use, dose reduction, intermittent use, or discontinuation. Clinician-directed treatment adjustment, planned cessation, or medication replacement was not classified as low adherence. When the available record was insufficient to determine implementation, adherence was coded as missing. Detailed operational definitions and exclusion rules for adherence classification are provided in Supplementary Appendix S2. Because dose counts and prescribed-dose denominators were not consistently available, no numerical threshold was retrospectively imposed.

Text Processing and Barrier Framework

Free-text data were obtained from the narrative clinical notes generated during routine follow-up. Free-text records were normalized by handling null values, trimming spaces, and collapsing repeated whitespace. We separated adherence behavior from the reported reason for that behavior. The behavior layer included regular use, complete or temporary discontinuation, irregular or reduced use, device-use error, and unclear information. The barrier layer included five prespecified reasons for difficulty with maintenance inhaler use: (1) medication beliefs/perceived effectiveness, such as perceived lack of benefit, feeling treatment was unnecessary after improvement, or unwillingness to continue long-term treatment; (2) side-effect concerns; (3) medication access/refill burden, including supply, cost, refill, visit, or transport difficulty; (4) forgetfulness/routine disruption; and (5) inhaler technique barriers. Operational definitions and exclusion rules for the five barrier domains are provided in Supplementary Appendix S3.

The five barrier categories were prespecified before the association analyses based on commonly described dimensions of inhaler non-adherence in the COPD literature and their applicability to routine follow-up documentation. Expressions encountered in the clinical notes were then used to operationalize these predefined domains into transparent keyword and regular-expression rules. The categories were not selected according to their statistical associations with adherence in the present dataset. Manual validation therefore evaluated the performance of the extraction rules for these prespecified domains rather than establishing the conceptual validity of the barrier framework itself. Keyword and regular-expression rules included exclusions for clinician-directed treatment changes, negated statements, warnings or hypothetical statements, resolved device problems, and questions without evidence that a problem had occurred. A record could contain more than one barrier category. Stopping, reducing, or irregularly using treatment was treated as medication-taking behavior rather than automatically as a barrier. A barrier was assigned only when the narrative contained affirmative evidence of the corresponding reason. Low adherence without an explicit documented reason remained classified as low adherence but did not contribute to a specific barrier category.

For analysis, a barrier category was coded as present only when the evaluable note contained affirmative evidence satisfying the corresponding rule. A category coded as zero should therefore be interpreted as “not documented in that note”, rather than as confirmed absence of the barrier. Patient-wave records without evaluable follow-up text were not treated as barrier-negative. The adherence classification was obtained from the routine clinical follow-up assessment; the rule-based text algorithm was used to extract barrier categories and did not generate the adherence outcome. The five categories represented clinically distinct barrier domains and were not assumed to have equivalent severity or clinical importance. The number of distinct categories was therefore used only as a simple exploratory summary of the diversity of documented barrier domains. The count was not intended to constitute a validated severity scale, weighted clinical score, or biological dose, and was not interpreted as evidence of a dose–response relationship.

Manual Validation

A simple random sample without replacement of 150 patient-wave records was drawn from records with evaluable follow-up text. Two reviewers independently labeled each of the five barrier categories and were blinded to one another’s labels. Because medication-taking behavior could be evident within the source free-text note itself, complete blinding to adherence-related information could not be guaranteed. Disagreements were adjudicated by a third reviewer.

Against the adjudicated labels, we calculated sensitivity, specificity, positive and negative predictive values, F1 score, and Cohen’s κ. Wilson 95% CIs were used for sensitivity, specificity, and predictive values. F1 and κ intervals used 10,000 fixed-seed record-level percentile bootstrap resamples. This validation addressed the five barrier categories only, not the behavior layer or adherence classification.

Statistical Analysis

Descriptive and Exploratory Analyses

Continuous variables are summarized as mean (standard deviation [SD]) and categorical variables as n (%). Wave-specific adherence proportions used participants with a nonmissing adherence assessment as the denominator. Barrier proportions used records with evaluable text. Adherence-state transitions were described among participants with observed adherence at both adjacent waves. Exploratory same-wave barrier comparisons used Fisher’s exact tests with Holm adjustment across 15 comparisons. For descriptive Wave 1 comparisons, age was compared using the Wilcoxon rank-sum test, and categorical variables were compared using Fisher’s exact test because the low-adherence group was very small.

Primary Temporally Ordered Analysis

For Wave 3, the primary exposure was the presence of any barrier documented during Waves 1–2 vs none. This binary exposure was selected because the five barrier categories represented heterogeneous clinical domains and because only two participants had two or more distinct prior barrier categories, limiting the ability to support a graded primary analysis. Firth penalized logistic regression was used with low adherence at Wave 3 as the outcome and standardized age and sex as covariates.²⁰ As a secondary exploratory analysis, we also examined the number of distinct prior barrier categories documented during Waves 1–2. This unweighted count was used as a transparent summary of the diversity of documented barrier domains. No validated clinical weights were available for the five categories, and the sparse category frequencies did not support reliable data-derived weighting. The count was therefore not interpreted as a severity score or as evidence of a linear dose–response relationship.

Secondary and Supportive Analyses

Adjacent-wave Firth models were supportive temporally ordered analyses. The outcome was high adherence at the next wave; predictors included the five barriers recorded at the current wave where estimable, current adherence, standardized age, and sex. The GEE analysis was secondary and exploratory because barrier and adherence information could originate from the same wave and note. A population-averaged logistic model used an exchangeable working correlation, participant identifier as the cluster, and high adherence as the outcome.²¹ Covariates were the four estimable same-wave barriers, age standardized by its sample SD, sex, and wave. Medication access/refill burden was omitted because all 10 exposed observations had low adherence, causing complete separation. Independence and first-order autoregressive working correlations were sensitivity analyses; QIC was reported descriptively.

Missing-Outcome Sensitivity Analysis

To assess missing Wave 3 outcomes, we fitted a logistic observation model for Wave 3 adherence observed (1) vs missing (0). Candidate variables were standardized age; sex; Wave 1 and Wave 2 adherence; missing indicators for age, sex, and both earlier adherence measures; and the Waves 1–2 barrier count. Age and sex had no missing values, so their missing indicators had no variation and were removed. Age would have been median-imputed and standardized; no age imputation occurred. Sex would have been mode-imputed; no sex imputation occurred. Missing Wave 1 and Wave 2 adherence values (14 and 31, respectively) were mode-imputed as high adherence, with separate indicators coded 1 for originally missing and 0 for observed. To prevent numerically extreme inverse-probability weights if estimated observation probabilities approached 0 or 1, predicted probabilities were prespecified to be bounded at 0.01–0.99 before weight construction. In the present dataset, all fitted probabilities (0.330–0.977) were already within these bounds, so this safeguard did not alter the fitted probabilities. Stabilized weights were subsequently truncated at the 1^st and 99^th percentiles. Stabilized weights equaled the marginal Wave 3 observation probability divided by each predicted probability and were truncated at the 1^st and 99^th percentiles.^22,23 Weighted quasibinomial logistic models with robust standard errors examined low adherence at Wave 3 in relation to prior barrier count or any prior barrier, adjusted for age and sex. This analysis relies on positivity and conditional missing-at-random assumptions and cannot address unmeasured determinants of missingness.²⁴

Results

Cohort and Baseline Characteristics

The cohort included 278 participants. At Wave 1, adherence was evaluable in 264 participants; their mean age was 67.9 (SD 8.4) years and 237 (89.8%) were male. High adherence was recorded in 259 participants and low adherence in five. Because the low-adherence baseline group was very small, baseline comparisons are descriptive and unstable (Table 1).

Table 1.

Wave 1 Characteristics Among Participants with Evaluable Adherence

Characteristic	Overall (n=264)	High Adherence (n=259)	Low Adherence (n=5)	P value
Age, years	67.9 ± 8.4	67.9 ± 8.4	71.2 ± 7.9	0.405
Male sex	237 (89.8%)	232 (89.6%)	5 (100.0%)	1.000
Stopped medication behavior	3 (1.1%)	1 (0.4%)	2 (40.0%)	<0.001
Irregular/underdosed behavior	3 (1.1%)	1 (0.4%)	2 (40.0%)	<0.001
Medication belief/perceived effectiveness	2 (0.8%)	1 (0.4%)	1 (20.0%)	0.038
Side-effect concern	7 (2.7%)	5 (1.9%)	2 (40.0%)	0.006
Access/refill burden	0	0	0	—
Forgetfulness/routine disruption	2 (0.8%)	1 (0.4%)	1 (20.0%)	0.038
Inhaler technique barrier	6 (2.3%)	5 (1.9%)	1 (20.0%)	0.109

Notes: Values are mean ± SD or n (%). Age was compared using the Wilcoxon rank-sum test (W=506.5). Categorical variables were compared using Fisher’s exact test because the low-adherence group contained only five participants. Fisher’s exact test does not provide a single conventional test statistic analogous to a Pearson χ² statistic; therefore, exact P values are reported. No comparison was performed for access/refill burden because there was no variation at Wave 1.

Manual Validation of Barrier Extraction

In the 150-record validation sample, category-specific sensitivity ranged from 75.0% to 93.8%, specificity from 97.8% to 100.0%, F1 scores from 0.750 to 0.968, and Cohen’s κ values from 0.736 to 0.964. Performance estimates were least precise for the access/refill category because positive reference records were uncommon. These findings support use of the rule-based categories for aggregate analyses but do not imply that an absent rule flag confirms absence of a barrier.

Adherence Behavior, Barriers, and State Transitions

Adherence was evaluable in 264, 247, and 236 participants at Waves 1, 2, and 3. High adherence was recorded in 259/264 (98.1%), 222/247 (89.9%), and 189/236 (80.1%), respectively (Table 2). The remaining 14, 31, and 42 outcomes were missing or not assessed; absence of evaluable text accounted for 8, 19, and 35 of these outcomes, and text indicating hospitalization or acute illness accounted for 2, 2, and 3. Two hundred participants (71.9%) had observed adherence at all three waves; 36 (12.9%) had only the Wave 3 outcome missing.

Table 2.

Wave-Specific Adherence and Recorded Barriers

Measure	Wave 1	Wave 2	Wave 3
Evaluable follow-up text, n	270	259	243
Evaluable adherence, n	264	247	236
High adherence, n/N (%)	259/264 (98.1)	222/247 (89.9)	189/236 (80.1)
Low adherence, n/N (%)	5/264 (1.9)	25/247 (10.1)	47/236 (19.9)
Belief/perceived effectiveness, n/N (%)	2/270 (0.7)	12/259 (4.6)	22/243 (9.1)
Side-effect concern, n/N (%)	8/270 (3.0)	9/259 (3.5)	4/243 (1.6)
Access/refill burden, n/N (%)	0/270 (0.0)	2/259 (0.8)	8/243 (3.3)
Forgetfulness/routine disruption, n/N (%)	2/270 (0.7)	8/259 (3.1)	5/243 (2.1)
Inhaler technique barrier, n/N (%)	6/270 (2.2)	7/259 (2.7)	3/243 (1.2)

Notes: Adherence percentages use evaluable adherence as denominator; barrier percentages use evaluable follow-up text. These are descriptive wave-specific proportions with changing denominators.

Among participants with adjacent-wave assessments, 211/231 (91.3%) with high adherence at Wave 1 remained highly adherent at Wave 2, while 20 (8.7%) changed to low adherence. From Wave 2 to Wave 3, 166/195 (85.1%) with high adherence remained high and 29 (14.9%) changed to low. Of 15 participants with low adherence at Wave 2 and observed Wave 3 adherence, 11 (73.3%) remained low and four (26.7%) changed to high.

Across any wave, medication beliefs/perceived effectiveness were recorded for 35/278 participants (12.6%), side-effect concerns for 19 (6.8%), forgetfulness/routine disruption for 13 (4.7%), inhaler technique barriers for 13 (4.7%), and access/refill burden for 10 (3.6%) (Figure 1). The wave-specific distribution of the five documented barrier categories is shown in Figure 2A. Same-wave barrier proportions by adherence status are reported in Supplementary Table S1 and Supplementary Figure S2.

Figure 1.

Documented barriers to maintenance inhaler adherence across follow-up. Bars show the proportion of participants with each barrier documented at least once across the three follow-up waves. Barrier categories were non-mutually exclusive.

Figure 2.

Longitudinal patterns of adherence barriers and Wave 3 low adherence by prior barrier status. (A) Wave-specific proportions of the five documented barrier categories. (B) Proportion of participants with low adherence at Wave 3 according to whether any barrier was documented during waves 1–2. Error bars indicate 95% confidence intervals. More than one barrier could be documented in the same record.

Primary Temporally Ordered Analysis

Among 236 participants with observed Wave 3 adherence, low adherence was recorded in 31/199 (15.6%) participants with no barrier documented during Waves 1–2 and in 16/37 (43.2%) participants with at least one prior barrier (Figure 2B). In the primary adjusted Firth model, any prior barrier vs none was associated with higher odds of low adherence at Wave 3 (OR 4.29, 95% CI 2.00–9.16). In the secondary exploratory count analysis, low adherence occurred in 31/199 (15.6%), 15/35 (42.9%), and 1/2 (50.0%) participants with 0, 1, and ≥2 distinct prior barrier categories, respectively (Supplementary Figure S3). The estimated OR per additional category was 3.67 (95% CI 1.81–7.58). Although the Cochran–Armitage test indicated an unadjusted ordered pattern (P<0.001), only two participants were in the ≥2-category group; the count analysis was therefore not interpreted as evidence of a stable graded or dose–response relationship.

Secondary Contemporaneous GEE Analysis

The exchangeable GEE included 747 patient-wave observations from 278 participants and returned finite estimates and standard errors. Each of the four estimable same-wave barrier categories was associated with lower odds of high adherence: beliefs/perceived effectiveness OR 0.013 (95% CI 0.004–0.039), side-effect concern OR 0.080 (0.023–0.286), forgetfulness/routine disruption OR 0.036 (0.009–0.136), and inhaler technique barrier OR 0.055 (0.012–0.255) (Table 3). Medication access/refill burden was not estimated because all 10 exposed observations had low adherence. Compared with Wave 1, high adherence was less likely to be recorded at Wave 2 (OR 0.208, 95% CI 0.070–0.616) and Wave 3 (OR 0.063, 95% CI 0.023–0.174). Age and sex were not statistically associated with the outcome. Because barrier and adherence classifications could arise from the same wave and note, these estimates characterize contemporaneous co-occurrence rather than temporal effects.

Table 3.

Exchangeable GEE for High Adherence

Variable	OR	95% CI	P value
Medication belief/perceived effectiveness	0.013	0.004–0.039	<0.001
Side-effect concern	0.080	0.023–0.286	<0.001
Forgetfulness/routine disruption	0.036	0.009–0.136	<0.001
Inhaler technique barrier	0.055	0.012–0.255	<0.001
Age, per 1 SD	1.076	0.750–1.544	0.690
Male sex	0.920	0.249–3.398	0.900
Wave 2 vs Wave 1	0.208	0.070–0.616	0.005
Wave 3 vs Wave 1	0.063	0.023–0.174	<0.001
Access/refill burden	Not estimable	Complete separation	—

Notes: N=747 observations from 278 participants. ORs are mutually adjusted for listed estimable barriers, standardized age, sex, and wave. The model is contemporaneous and should not be interpreted causally.

Additional Temporally Ordered and Sensitivity Analyses

In the Wave 1→2 Firth model (n=236), current high adherence was not statistically associated with next-wave high adherence (OR 4.35, 95% CI 0.53–33.59), and all individual barrier CIs were wide. In the Wave 2→3 model (n=210), current high adherence was associated with next-wave high adherence (OR 19.55, 95% CI 3.03–284.84), while no individual barrier reached P<0.05; the side-effect estimate was OR 0.225 (95% CI 0.042–1.370) (Table 4). These results support persistence of adherence status over the second interval but leave barrier-specific temporal associations uncertain.

Table 4.

Selected Temporally Ordered and Missing-Outcome Sensitivity Analyses

Analysis	Exposure	OR	95% CI	P value	N
Primary Firth model	Any prior barrier vs none	4.29	2.00–9.16	<0.001	236
Secondary exploratory Firth model	Per additional barrier category	3.67	1.81–7.58	<0.001	236
Truncated stabilized IPW	Any prior barrier vs none	4.81	2.20–10.50	<0.001	236
Secondary exploratory IPW analysis	Per additional barrier category	4.15	1.93–8.89	<0.001	236
Lagged Firth, Wave 1→2	Current high adherence	4.35	0.53–33.59	0.164	236
Lagged Firth, Wave 2→3	Current high adherence	19.55	3.03–284.84	0.001	210

Notes: The primary exposure was any barrier documented during Waves 1–2 vs none. The number of distinct prior barrier categories was retained as a secondary exploratory exposure and should not be interpreted as a validated severity scale or as evidence of a linear dose–response relationship. Firth prior-barrier models and weighted models were adjusted for standardized age and sex.

The independence and first-order autoregressive GEE models converged and produced similar directions and magnitudes; QIC values were 310.48, 309.80, and 309.91 for independence, first-order autoregressive, and exchangeable structures, respectively (Supplementary Table S2). Wave 3 adherence was observed in 236/278 (84.9%). The observation model converged; male sex (OR 3.27, 95% CI 1.26–8.06) and high adherence at Wave 2 (OR 5.35, 95% CI 1.56–19.22) were associated with higher odds that Wave 3 adherence was observed. Predicted observation probabilities ranged from 0.330 to 0.977. Truncated stabilized weights ranged from 0.869 to 1.476, had coefficient of variation 0.128, and yielded an effective sample size of 232.2. The weighted association for any prior barrier vs none was OR 4.81 (95% CI 2.20–10.50), while the secondary exploratory estimate per additional prior barrier category was OR 4.15 (95% CI 1.93–8.89). Similarity to the unweighted Firth estimates indicates robustness to this specified weighting model, conditional on its assumptions, but does not eliminate bias from unmeasured causes of missingness (Table 4). In individual-barrier Firth analyses, all five barriers were associated with substantially lower odds of concurrent high adherence at Wave 3 (ORs 0.006–0.053; all P≤0.008), whereas none was statistically significant in either lagged model (Supplementary Table S3).

Discussion

In this three-wave cohort, high adherence was recorded less frequently at successive approximately monthly follow-ups, and the composition of documented barriers changed across waves. Medication beliefs/perceived effectiveness were the most frequently documented category overall and were recorded more often at later waves, while access/refill burden was absent at Wave 1 and recorded in eight Wave 3 notes. These descriptive findings show that the information elicited during follow-up was not static; they do not by themselves demonstrate within-person deterioration or the emergence of a causal barrier.

The primary temporally ordered analysis found that any barrier documented during Waves 1–2, compared with none, was associated with low adherence at Wave 3 in both unweighted and weighted analyses. A secondary exploratory analysis using the number of distinct prior barrier categories yielded an estimate in the same direction, and the unadjusted Cochran–Armitage test also indicated a statistically significant ordered trend (P<0.001). However, this pattern was driven largely by the difference between participants with no recorded barrier and those with at least one barrier, while the group with at least two barriers contained only two participants. The binary distinction between any documented prior barrier and none was therefore selected as the primary summary because the five barrier categories represented clinically heterogeneous domains and the higher-count groups were sparse. The exploratory category count was intended only to summarize the diversity of documented barrier domains, not their severity or clinical importance. Because the categories were unweighted and only two participants had two or more prior categories, the count estimate should not be interpreted as a validated cumulative score or a dose–response relationship. This binary distinction is therefore the more readily interpretable summary in the present dataset, although it also requires external validation. Wave 2 adherence status was also associated with Wave 3 status, although the wide CI reflected sparse transitions and limited precision. None of the individual barrier categories was statistically associated with next-wave adherence in the fully adjusted lagged models. The very large same-wave GEE estimates probably reflect, at least partly, how a barrier and low adherence were documented together in the same note. They should not be interpreted as independent causal effects or used to rank barriers by importance.

Our findings are broadly consistent with literature describing beliefs, adverse-effect concerns, practical access problems, forgetfulness, and device technique as relevant dimensions of inhaler adherence.^4,6–13 They also extend routine-care studies by explicitly separating reported behavior from reasons and by distinguishing contemporaneous from lagged analyses. Transparent rules allowed source-level audit, while manual validation suggested high specificity across categories. Sensitivity was lower for beliefs and device technique, and the access category had few positive reference records; therefore an absent rule flag cannot be assumed to mean that a barrier was absent.

The patient-preference implication is modest but clinically relevant. A brief, structured check of maintenance inhaler use can be supplemented by an open question about why use differed from the prescription. When a concern is documented, confirmation with the patient or caregiver is essential before acting. Pharmacist- and clinician-led interventions can improve inhaler adherence and technique in some settings,²⁵ but this observational study did not evaluate an intervention and cannot determine which response would be effective for a particular barrier.

Strengths and Limitations

Strengths include repeated assessments at prespecified intervals, an explicit binary definition of reported maintenance inhaler use, separation of behavior from barrier categories, independent manual review by two reviewers with third-reviewer adjudication, and retention of model and weighting audit outputs. Temporally ordered models and observation weighting addressed two important interpretive concerns, although neither converts the study into a causal design.

Several limitations require emphasis. First, the study was conducted at one hospital and included predominantly men, limiting generalizability. Second, adherence was reported by patients or caregivers and was not corroborated by dose counters, electronic monitors, refill records, or drug exposure measures; social desirability and recall error are possible. Third, low adherence was rare at Wave 1 and several barrier categories were uncommon. Sparse exposure-outcome combinations produced separation and very wide confidence intervals in some models, particularly the barrier-specific lagged analyses. The highest prior-barrier-count category contained only two participants. These results should therefore be interpreted as imprecise and hypothesis-generating. Fourth, documentation intensity and wording may have changed over follow-up, and an unrecorded barrier may have been present. Fifth, the validation sample contained few positives for some categories and did not validate the behavior layer or adherence outcome. Sixth, the observation-weighting analysis depends on a correctly specified model, adequate positivity, and missingness conditional on measured variables; unmeasured reasons for non-observation remain possible. Seventh, available models did not include education, health literacy, socioeconomic status, smoking, regimen complexity, device type, exacerbations, comorbidity, or clinician communication. Residual confounding by these and other unmeasured patient- or treatment-level factors therefore cannot be excluded. Finally, different wave-specific denominators mean that descriptive proportions should not be interpreted as an average within-person change.

Conclusion

In this single-center COPD cohort, high maintenance inhaler adherence was recorded less frequently at later follow-ups, and the presence of any barrier documented during the first two waves was associated with low adherence at Wave 3. Adherence status showed persistence over the second interval, whereas individual barrier-specific lagged associations remained uncertain. Routine follow-up text may complement structured inquiry by preserving patient-reported context, but these findings do not establish causation or individual prediction. Future multicenter studies incorporating objective adherence measures and richer clinical covariates are needed to validate these associations and to determine whether barrier-informed follow-up can improve adherence.

Funding Statement

This work was supported by the Haiyan County Health Bureau (2024SY02). The funding organization was not involved in the study design, data collection, statistical analysis, result interpretation, or manuscript preparation.

Use of Generative Artificial Intelligence

No generative artificial intelligence tool was used to create or modify the study figures.

Abbreviations

CI, confidence interval; COPD, chronic obstructive pulmonary disease; GEE, generalized estimating equations; IPW, inverse probability weighting; OR, odds ratio; QIC, quasi-likelihood under the independence model criterion; SD, standard deviation.

Data Sharing Statement

The data that support the findings of this study are not publicly available due to privacy and ethical restrictions but are available from the corresponding author (Min-feng Zhou) upon reasonable request and with permission from the relevant ethics committee.

Ethics Approval and Consent to Participate

The study was conducted in accordance with the Declaration of Helsinki. The Ethics Committee of Haiyan People’s Hospital approved the study (approval number: 2024-33) and waived the requirement for informed consent. All procedures adhered to the principles of the Declaration of Helsinki.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.