Longitudinal Monitoring and Predictive Modeling of Medication Adherence in Ischemic Stroke Patients

Summary

This protocol describes structured telephone assessment of self-reported medication-taking at 30, 90, and 180 days after discharge and the internal development of an early follow-up prediction model for ischemic stroke survivors. The protocol separates self-reported ingestion from pharmacy-based PDC and uses predefined demographic and day-30 psychosocial predictors.

Abstract

Medication adherence is central to secondary prevention after ischemic stroke, but self-reported adherence may change during follow-up. In this single-center longitudinal framework, 210 participants were enrolled, and 191 completed the 180-day assessment. Medication-taking was assessed using a structured telephone interview for days 1–30, 31–90, and 91–180, and was classified as good when the exposure-day-weighted composite adherence percentage was at least 80%. The Chinese BMQ-Specific and Family APGAR were administered at day 30, day 90, and day 180. A four-predictor logistic model used residence, sex, day-30 BMQ-NCD, and day-30 Family APGAR and was internally validated with 1,000 bootstrap resamples. Among 191 complete cases, good adherence was 84.3% (95% CI, 78.3%–89.1%) at day 30, 77.0% (95% CI, 70.3%–82.7%) at day 90, and 72.8% (95% CI, 65.9%–79.0%) at day 180. Urban residence, male sex, day-30 BMQ-NCD, and day-30 Family APGAR score were associated with day-180 adherence. The apparent AUC was 0.879 (95% CI, 0.821–0.928), and the optimism-corrected AUC was 0.867. The protocol provides a reproducible framework for longitudinal assessment of self-reported adherence and an internally validated day-30 prediction model. External validation and objective adherence measures are required before clinical implementation.

Introduction

Stroke remains a leading cause of long-term disability and mortality worldwide, necessitating a refined clinical definition that accounts for both pathological and imaging evidence¹. Despite advancements in acute interventions, the global burden of stroke continues to escalate, with ischemic stroke accounting for the vast majority of cases². This trend is particularly pronounced in China, where the prevalence of vascular risk factors has reached critical levels, placing an immense strain on the healthcare system³. Evidence suggests that survivors of a first incident stroke face a high risk of recurrent vascular events, with community-based studies highlighting the persistent vulnerability of this population⁴. While managing the acute phase presents significant challenges⁵, the long-term survival and life expectancy for these patients are predominantly dictated by the efficacy of secondary prevention strategies^6,7.

The cornerstone of secondary prevention involves rigorous pharmacological management, including blood pressure lowering and antiplatelet therapy^8,9,10. International guidelines from the American Heart Association (AHA), the European Stroke Organisation (ESO), and the Canadian Stroke Best Practice Recommendations emphasize that adherence to these regimens is essential to prevent recurrence^11,12,13,14. Nevertheless, the clinical utility of these interventions is frequently undermined by poor medication adherence^15,16. Medication adherence is a multi-dimensional construct that includes both initiation and persistence^15,17. Recent registry data indicate that greater adherence to secondary prevention medications significantly improves survival and reduces recurrence rates^17,18. For example, statin adherence has been independently associated with reduced recurrent risk¹⁹.

Despite these benefits, longitudinal studies observe a significant “temporal decay” in adherence rates during the first-year post-discharge²⁰. While demographic factors such as sex differences and socioeconomic status influence persistence^21,22, there is an increasing recognition of the role of psychosocial and cognitive factors in shaping patient behavior^23,24. The Beliefs about Medicines Questionnaire (BMQ) has emerged as a validated tool to assess the cognitive representation of medication through the “Necessity-Concerns” framework²⁵. Systematic reviews have demonstrated that BMQ scores are robust predictors of adherence across various chronic conditions in China²⁶. Similarly, the social environment, particularly family functioning, plays a pivotal role. The Family APGAR Index provides a reliable measure of perceived support, which is critical for patients navigating the complexities of post-stroke life^27,28.

While the NIHSS and modified Rankin Scale describe neurological severity and functional outcome^29,30, they do not directly measure medication beliefs or perceived family functioning³¹. Mobile health and WeChat-based services have been studied as adherence supports^32,33. The present protocol, therefore, has two aims: to define medication-taking assessment at day 30, day 90, and day 180 using consistent terminology, and to develop an interpretable day-180 adherence model using predefined demographic and psychosocial variables collected during the study.

Protocol

The protocol was approved by the Ethics Committee on Biomedical Research, West China Hospital of Sichuan University (Approval No. 2021 Review (1303); approval date, 8 November 2021). All reported procedures were conducted in accordance with the committee's requirements and the approved protocol. The complete analytic dataset is provided in Supplementary File 1.

1. Patient screening and baseline enrollment

1. Identify potential participants from the inpatient records of the Department of Neurology. Select patients diagnosed with acute ischemic stroke via magnetic resonance imaging (MRI) or computed tomography (CT). Record the total number of initially screened individuals to establish the baseline pool for the participant flow diagram.
2. Screen patients against the prespecified inclusion criteria. Include adults with imaging-confirmed acute ischemic stroke who can complete the interview themselves or through a legally authorized representative.
3. Apply exclusion criteria. Exclude patients with a history of severe mental illness, cognitive impairment (determined by clinical records), or a life expectancy of less than six months. Document the exact reasons and corresponding numbers for any exclusions at this stage (e.g., life expectancy limitations or refusal to provide consent).
4. Before enrollment, obtain written informed consent from the patient or, when the patient cannot provide consent, from a legally authorized representative, using the committee-approved informed consent form. Explain the longitudinal study design, telephone assessments, and 180-day follow-up schedule before signature. Record the consenting person's identity and role, the consent date, and the study staff member who obtained consent.
5. Collect baseline demographic and clinical data within 48 h of admission. Record age, sex, residence, educational level, and monthly income using a standardized case report form.
6. At discharge, have a trained stroke clinician administer the validated Chinese NIHSS using its standard 11 scored items and total range of 0–42²⁹. Record the total score, assessor role, and training status, assessment date, and length of hospital stay.
7. Administer the authorized Chinese BMQ-Specific by a structured face-to-face interview before discharge and by telephone at day 30, day 90 and day 180^25,34.
   NOTE: The instrument contains five Necessity items and five Concerns items. Score each item from 1 (strongly disagree) to 5 (strongly agree), and sum the five items in each subscale; Necessity and Concerns totals, therefore each range from 5 to 25.
8. Calculate the necessity-concerns differential (BMQ-NCD) as Necessity minus Concerns (range, -20 to 20). Higher necessity and higher BMQ-NCD values indicate stronger perceived need relative to concerns, whereas a higher Concerns value indicates greater apprehension.
   NOTE: Require all five responses for each subscale; otherwise, code that subscale and the BMQ-NCD as missing. Use day 30 BMQ-NCD as the continuous model predictor. At the same time, administer the Chinese Family APGAR²⁷.
9. Score Adaptation, Partnership, Growth, Affection, and Resolve as 0 (hardly ever), 1 (some of the time), or 2 (almost always), and sum the five items to an integer total of 0-10; higher scores indicate better perceived family functioning. If any Family APGAR item is missing, code the total as missing. Record the instrument versions, authorization, administrator, respondent, date, item responses, subscale totals, and total scores on the case-report form.

2. Discharge preparation and patient education

1. Conduct a standardized medication counseling session before discharge. Provide a written list of prescribed secondary prevention medications, including antiplatelets, statins, and antihypertensives.
2. Instruct the patient and their primary caregiver on the importance of medication adherence. Demonstrate how to record medication intake if the patient intends to use a pillbox or diary.
3. Verify the patient’s primary telephone contact number. Establish a preferred time window for future follow-up calls to minimize attrition and ensure a high response rate.

3. Post-discharge longitudinal follow-up (Day 30, Day 90, and Day 180)

1. Conduct the first structured telephone interview at day 30 after discharge (target window, plus or minus 3 days). For every secondary-prevention medication class prescribed during days 1–30, record the class name, prescribed daily dose frequency, number of days exposed, and the self-reported number of class-adherent days during days 1–30.
   1. Record the respondent (patient or primary medication-managing caregiver), contact date, regimen changes, temporary physician-directed interruptions, and reasons for noncompletion. At the same interview, administer the BMQ-Specific and Family APGAR and record the day-30 scores
2. Conduct the second interview at day 90 (target window, plus or minus 7 days). Repeat the class-specific medication questions for the nonoverlapping days 31–90 interval and administer the BMQ and Family APGAR. Do not ask the participant to reconstruct days 1–30 again; retain the day-30 record as the source for that interval.
3. Conduct the final interview at day 180 (target window, plus or minus 7 days). Repeat the class-specific medication questions for the nonoverlapping days 91–180 interval, administer the BMQ and Family APGAR, and record the day-180 mRS when available.
4. Use the same structured interview form and respondent rule at all three follow-ups. When a patient cannot respond because of cognitive, language, or physical impairment, use the caregiver who manages the medication regimen and record that substitution. Do not claim pill-count, diary, pharmacy, or interviewer-blinding verification unless it is documented in the source records.
5. Record outpatient visits for descriptive follow-up only. Do not include any post-discharge utilization variable in a prediction model intended for use after the day-30 assessment.
6. Define baseline as the inpatient/discharge assessment and label adherence assessments by calendar time (day 30, day 90, and day 180), rather than T1, T2, and T3. This prevents the previous conflict in which T1 referred both to baseline and to day 30.

4. Quantification of self-reported medication adherence

1. Do not use the term PDC unless adherence is derived from dispensing or refill records. Treat antithrombotic therapy (antiplatelet or oral anticoagulant, as indicated), lipid-lowering therapy, antihypertensive therapy, and glucose-lowering therapy as separate secondary-prevention classes only when prescribed to that participant.
   1. For each prescribed class, define a class-adherent day as a day on which every scheduled dose in that class was reportedly taken; omission of any scheduled dose makes that class-day nonadherent. Physician-directed discontinuations, substitutions, and temporary holds are not counted as expected exposure days after the documented effective date.
   2. Calculate the interval composite self-reported adherence percentage as 100 multiplied by the sum of class-adherent days across prescribed classes divided by the sum of expected class-exposure days across those classes. This exposure-day-weighted calculation accommodates different prescription durations and is equivalent to the mean of class-specific proportions only when every class is prescribed for the same number of days. Use days 1–30 for the day-30 value, days 31–90 for the day-90 value, and days 91–180 for the day-180 value.
2. At each follow-up, classify good adherence as a self-reported proportion of adherent days of at least 80% and poor adherence as less than 80%. Treat missing adherence assessments as missing; do not assign them to the poor-adherence category.

5. Data integration, longitudinal analysis, and predictive modeling

1. Perform the reproducible analysis using the Python programming language (version 3.13.5) with libraries for data manipulation, numerical computation, statistical analysis, and visualization on a 64-bit Windows operating system. Set the random seed to 71636, and code missing values as NA. If the final team instead uses R, replace this section with the exact executed R version, package versions, functions, and script; do not report an unexecuted software workflow.
2. Summarize good-adherence proportions at day 30, day 90, and day 180 with two-sided 95% Clopper-Pearson exact confidence intervals. Test the overall paired change in binary adherence with Cochran's Q test, followed by exact paired McNemar tests with Bonferroni correction.
3. Analyze BMQ-NCD and Family APGAR across day 30, day 90, and day 180 among complete cases using the Friedman test. When the omnibus test is significant, conduct paired Wilcoxon signed-rank comparisons with Bonferroni correction. Report medians and interquartile ranges in addition to means and standard deviations. Calculate individual scale totals from item-level responses before analysis; do not analyze interpolated or manually smoothed questionnaire values.
4. Define the day-180 outcome as good adherence = 1 and poor adherence = 0. Refit a day-30 multivariable logistic regression using four prespecified predictors: residence (Urban = 1; Rural = 0), sex (Male = 1; Female = 0), day-30 BMQ-NCD (continuous), and day-30 Family APGAR score (continuous). Do not use outpatient follow-up visits. Avoid describing univariate screening or backward selection unless that procedure was actually executed and documented in code.
5. Specify the complete fitted equation: logit[P(good adherence at day 180)] = -10.3811 + 1.2921(Urban) - 1.5215(Male) + 0.5211(BMQ-NCD) + 0.6887(Family APGAR). Generate the nomogram directly from these unrounded coefficients and use reader-facing labels, reference categories, score ranges, and directionality.
6. Assess discrimination with the ROC AUC and a percentile-bootstrap 95% confidence interval. Assess overall prediction error with the Brier score and calibration with a calibration plot and calibration slope. Report apparent performance separately from optimism-corrected performance.
7. Perform 1,000 successful bootstrap resamples. In each resample, refit the complete four-predictor model and estimate optimism by comparing performance in the bootstrap sample with performance in the original sample. Subtract mean optimism from apparent performance.

6. Patient confidentiality and data safety

1. Ensure all patient-identifiable information is removed before analysis. Store the master linkage file in a password-protected, encrypted institutional system accessible only to authorized study personnel.
2. Assign a unique alphanumeric ID to each participant at enrollment. Use this ID for all subsequent follow-up forms and digital records.

Results

Implementation of the standardized protocol and cohort stratification

The participant-flow log reports that 216 individuals were assessed, 6 were excluded, and 210 were enrolled. Seventeen were lost before the day-90 assessment, and two additional participants were lost before day 180, leaving 191 complete cases for the longitudinal and modeling analyses (Supplementary Figure 1). The 19 noncompleters were documented in the screening and follow-up log for participant-flow reporting but were not included in the analytic dataset, and no outcome values were imputed. The analysis file, therefore, contains the prespecified complete-case cohort of 191 participants.

At day 180, 139 of 191 participants were classified as having good adherence and 52 as having poor adherence (Table 1). Urban residence and sex differed between adherence groups in univariate comparisons, whereas other baseline variables should be reported with their exact effect estimates and P-values rather than described collectively as balanced. The female poor-adherence cell contained five participants, so sex-related estimates require cautious interpretation.

Quantitative characterization of adherence decay and transitions

Among the 191 complete cases, good adherence was 84.3% (161/191; exact 95% CI, 78.3%–89.1%) for days 1–30, 77.0% (147/191; 70.3%–82.7%) for days 31–90, and 72.8% (139/191; 65.9%–79.0%) for days 91–180 (Figure 2). The overall paired difference was significant (Cochran Q = 31.0, df = 2, P < 0.001). The observed day-30/day-90/day-180 patterns were good/good/good (n = 138), good/good/poor (n = 8), good/poor/poor (n = 15), poor/good/good (n = 1), and poor/poor/poor (n = 29) (Figure 3). We refer to these descriptively as adherence-state transitions, without implying that a new validated behavioral construct has been established.

Sensitivity and longitudinal responsiveness of psychosocial metrics

Among complete cases, mean BMQ-NCD values were 13.06 (SD 2.60) at day 30, 12.06 (3.03) at day 90, and 10.88 (3.84) at day 180; mean Family APGAR values were 8.20 (1.28), 7.79 (1.45), and 7.23 (1.84), respectively. Both Friedman omnibus tests and Bonferroni-adjusted paired Wilcoxon tests yielded P < 0.001 (Table 2; Figure 4). Individual questionnaire totals should be regenerated directly from the item-level case-report forms before the final analysis is locked.

Parameter weighting and synthesis of the predictive tool

The day-30 prediction model retained four prespecified predictorsretained four prespecified predictors (Table 3): urban residence (adjusted OR, 3.64; 95% CI, 1.54–8.62; P = 0.0033), male sex (adjusted OR, 0.22; 95% CI, 0.06–0.76; P = 0.0174), day 30 BMQ-NCD per point (adjusted OR, 1.68; 95% CI, 1.39–2.05; P < 0.001), and day 30 Family APGAR score per point (adjusted OR, 1.99; 95% CI, 1.39–2.86; P < 0.001). The fitted intercept was -10.3811. Figure 5 displays the same coding and one-point scale used in Table 3, and Figure 6 maps the unrounded coefficients to the nomogram.

Methodological validation: Accuracy and calibration

The model's apparent AUC was 0.879 (bootstrap 95% CI, 0.821–0.928), and the optimism-corrected AUC after 1,000 successful bootstrap resamples was 0.867 (Figure 7). The apparent Brier score was 0.118, and the optimism-corrected calibration slope was 0.923. These values describe internal performance in the 191 complete cases and do not constitute external validation.

Figure 1: Standardized operational roadmap for post-stroke adherence monitoring. This flowchart illustrates the four-phase protocol implemented in the study: Phase 1 (Enrollment and baseline), Phase 2 (Acute care discharge), Phase 3 (Systematic longitudinal monitoring), and Phase 4 (Risk evaluation and analysis). This roadmap serves as the Standard Operating Procedure (SOP) for the proposed methodology. Please click here to view a larger version of this figure.

Figure 2: Longitudinal good-adherence proportions. The figure shows complete-case good-adherence proportions for days 1–30 at the day-30 assessment (84.3%), days 31–90 at the day-90 assessment (77.0%), and days 91–180 at the day-180 assessment (72.8%). Error bars are two-sided 95% Clopper-Pearson exact confidence intervals. Adherence was classified as good when the exposure-day-weighted composite self-reported adherence percentage was at least 80%. Please click here to view a larger version of this figure.

Figure 3: Individual adherence-state transitions. The alluvial diagram shows each complete-case participant's good or poor adherence classification at day 30, day 90, and day 180. Ribbon widths represent participant counts. The term adherence-state transition is descriptive and is not presented as a previously validated construct. Please click here to view a larger version of this figure.

Figure 4: Family APGAR and BMQ-NCD distributions across time. Boxes representing (A) family APGAR and (B) BMQ-NCD distributions represent the interquartile range, center lines the median, whiskers values within 1.5 times the interquartile range, and points beyond the whiskers individual observations. Overall comparisons use Friedman tests; paired comparisons use Wilcoxon signed-rank tests with Bonferroni correction. BMQ-NCD is the Necessity total minus the Concerns total. Please click here to view a larger version of this figure.

Figure 5: Adjusted associations in the day-30 prediction model. Points are adjusted ORs and horizontal bars are 95% CIs for urban versus rural residence, male versus female sex, and one-point increases in baseline day 30 BMQ-NCD and Family APGAR. The horizontal axis is logarithmic, and the vertical reference line denotes OR = 1. Please click here to view a larger version of this figure.

Figure 6: Nomogram for internally estimated probability of good adherence at Day 180. The nomogram uses: residence, sex, day 30 BMQ-NCD, and day 30 Family APGAR score. Predictor values map to points; total points map to estimated probability. Higher BMQ-NCD indicates that perceived necessity outweighs concerns, and higher Family APGAR indicates better perceived family functioning. The nomogram requires external validation before clinical use. Please click here to view a larger version of this figure.

Figure 7: Internal validation of the day-30 prediction model. (A) The ROC curve shows an apparent AUC of 0.879 (bootstrap 95% CI, 0.821-0.928) and an optimism-corrected AUC of 0.867. (B) The calibration panel compares predicted and observed probabilities. The ideal line denotes perfect calibration, the apparent curve describes the fitted sample, and the bias-corrected curve reflects 1,000-resample bootstrap optimism correction. Shaded or boundary lines, if displayed, represent 95% confidence limits; rug marks show the distribution of predicted probabilities. Please click here to view a larger version of this figure.

Characteristic	Overall (N = 191)	Poor adherence (N = 52)	Good adherence (N = 139)	P value
Age, years	59.2 +/- 13.1	59.6 +/- 14.2	59.1 +/- 12.7	0.797
Age group, n (%)				0.822
<65 years	120 (62.8%)	32 (61.5%)	88 (63.3%)
>=65 years	71 (37.2%)	20 (38.5%)	51 (36.7%)
Sex, n (%)				0.015
Female	41 (21.5%)	5 (9.6%)	36 (25.9%)
Male	150 (78.5%)	47 (90.4%)	103 (74.1%)
Residence, n (%)				<0.001
Rural	92 (48.2%)	36 (69.2%)	56 (40.3%)
Urban	99 (51.8%)	16 (30.8%)	83 (59.7%)
Education, n (%)				0.224
Primary school or below	59 (30.9%)	21 (40.4%)	38 (27.3%)
Middle school	50 (26.2%)	13 (25.0%)	37 (26.6%)
High school/technical	36 (18.8%)	10 (19.2%)	26 (18.7%)
College degree or above	46 (24.1%)	8 (15.4%)	38 (27.3%)
Marital status, n (%)				0.905
Married	175 (91.6%)	47 (90.4%)	128 (92.1%)
Unmarried	9 (4.7%)	3 (5.8%)	6 (4.3%)
Divorced/widowed	7 (3.7%)	2 (3.8%)	5 (3.6%)
Occupation, n (%)				0.138
Employed/professional	36 (18.8%)	9 (17.3%)	27 (19.4%)
Worker	25 (13.1%)	7 (13.5%)	18 (12.9%)
Farmer	39 (20.4%)	16 (30.8%)	23 (16.5%)
Retired	45 (23.6%)	7 (13.5%)	38 (27.3%)
Self-employed/other	46 (24.1%)	13 (25.0%)	33 (23.7%)
Monthly income, CNY, n (%)				0.208
<3000	39 (20.4%)	14 (26.9%)	25 (18.0%)
3000-4999	56 (29.3%)	17 (32.7%)	39 (28.1%)
>=5000	96 (50.3%)	21 (40.4%)	75 (54.0%)
NIHSS score at discharge	2.0 [0.0-4.5]	2.0 [0.8-5.0]	2.0 [0.0-4.0]	0.603
Length of hospital stay, days	8.0 [6.0-10.0]	7.0 [5.0-9.0]	8.0 [7.0-10.0]	0.021

Table 1: Characteristics of the study population for methodological validation (N = 191). This table summarizes the demographic and clinical profiles of the participants, stratified by their 180-day self-reported medication adherence status (Good vs. Poor). Categorical variables are presented as frequencies (n) and percentages (%). P-values are provided for descriptive between-group comparisons and were not used for predictor selection.Please click here to download this Table.

Measure	Day 30, mean +/- SD	Day 90, mean +/- SD	Day 180, mean +/- SD	Change (day 180 - day 30), mean +/- SD	Friedman chi-square (df = 2)	Global P value
BMQ-NCD total score	13.06 +/- 2.60	12.06 +/- 3.03	10.88 +/- 3.84	-2.18 +/- 1.90	382.0	<0.001
Family APGAR score	8.20 +/- 1.28	7.79 +/- 1.45	7.23 +/- 1.84	-0.97 +/- 0.91	357.5	<0.001

Table 2: Longitudinal BMQ-NCD and family APGAR scores. This table reports each scale at day 30, day 90, and day 180 as mean (SD) and median (IQR). Overall P-values are from Friedman tests; paired post-hoc P-values are from Wilcoxon signed-rank tests with Bonferroni correction. BMQ-NCD is the BMQ-Specific Necessity total minus the Concerns total (range, -20 to 20).

Predictor	β coefficient	Standard error	Wald z	P value	Adjusted OR (95% CI)	Coding / unit
Intercept	-10.3811	2.0697	-5.016	<0.001	—	Model intercept
Urban residence	1.2921	0.4396	2.94	0.0033	3.64 (1.54–8.62)	Urban vs rural (reference)
Male sex	-1.5215	0.6395	-2.379	0.0174	0.22 (0.06–0.76)	Male vs female (reference)
Baseline BMQ-NCD	0.5211	0.0995	5.238	<0.001	1.68 (1.39–2.05)	Per 1-point increase
Baseline Family APGAR	0.6887	0.1847	3.728	<0.001	1.99 (1.39–2.86)	Per 1-point increase

Table 3: Full multivariable logistic regression specification for good adherence at day 180 (N = 191). The table reports the intercept and regression coefficients, standard errors, Wald statistics, P-values, adjusted ORs, and 95% confidence intervals for the final multivariable logistic regression model. Good adherence is coded 1, and poor adherence is coded 0. Residence is Urban = 1 versus Rural = 0; sex is Male = 1 versus Female = 0; day 30 BMQ-NCD and day 30 Family APGAR are modeled per one-point increase.

Supplementary Figure 1: Participant flow and cohort selection for the longitudinal medication adherence study.Please click here to download this file.

Supplementary File 1: The complete analytic dataset.Please click here to download this file.

Discussion

This study describes a standardized longitudinal telephone protocol and internally develops a day-30 prediction model for day-180 self-reported medication adherence after ischemic stroke. Three findings are central. First, the proportion classified as having good adherence decreased from 84.3% at day 30 to 77.0% at day 90 and 72.8% at day 180. Second, BMQ-NCD and Family APGAR scores also decreased across the scheduled assessments. Third, urban residence, sex, day 30 BMQ-NCD, and day 30 Family APGAR were associated with day-180 adherence in the prespecified day-30 prediction model. These findings support repeated post-discharge assessment while remaining descriptive and internally predictive rather than causal.

The reduction in good-adherence prevalence is consistent with prior stroke research showing that persistence with secondary-prevention medication may decline after discharge^17,18,19,20. Most participants remained in the good-adherence category, whereas the largest changing group moved from good to poor adherence. This pattern suggests that a satisfactory early interview should not be treated as proof of sustained medication-taking. Clinically, day-90 reassessment may identify patients whose behavior changed after the immediate recovery period, and the day-180 interview provides a later opportunity to review regimen complexity, adverse effects, access barriers, and caregiver involvement.

An important methodological contribution is the separation of the three nonoverlapping recall windows: days 1–30, days 31–90, and days 91–180. This approach avoids repeatedly asking participants to reconstruct the entire post-discharge period and links each estimate to a defined interval. The exposure-day-weighted composite also accommodates medication classes with different prescribed durations. Nevertheless, it remains a self-reported proportion of class-adherent days, not pharmacy-derived PDC. Recall error and social-desirability bias may therefore overestimate ingestion. The observed percentages should be interpreted as structured interview estimates rather than objective confirmation that medication was dispensed or taken.

The longitudinal BMQ-NCD findings provide a possible psychosocial context for the adherence pattern. BMQ-NCD is the Necessity score minus the Concerns score and therefore represents the relative balance between perceived need and apprehension, rather than a global BMQ total^25,26,34. A lower value may reflect weaker necessity beliefs, stronger concerns, or both. Repeated administration can help clinicians identify which component changed and tailor counselling accordingly, but the present analyses cannot determine whether changes in beliefs preceded changes in adherence. Questionnaire responses may also be influenced by health status and treatment experience. Item-level records and the exact Chinese instrument should be verified before drawing mechanistic conclusions.

Family APGAR similarly measures perceived family functioning rather than the amount of objectively observed caregiving^27,28. During stroke recovery, changes in disability, activities of daily living, psychological distress, caregiver burden, household roles, and culturally shaped expectations may alter how support is experienced^{35,36,37,38,39,40}. The decline in Family APGAR should consequently be described as an association within this cohort, not evidence that family support inevitably deteriorates or causes nonadherence. The result nevertheless supports caregiver involvement in follow-up and repeated assessment of practical support and treatment roles.

The revised prediction model is intended for use after the day-30 follow-up because two psychosocial predictors were collected at that assessment. Post-discharge outpatient visits occurring after day 30 were excluded to reduce temporal ambiguity. Higher day-30 BMQ-NCD and Family APGAR scores were associated with greater odds of good adherence at day 180, supporting the potential value of early psychosocial reassessment. Post-discharge outpatient visits were excluded because they are contemporaneous with the outcome window and could introduce temporal ambiguity. Urban residence may represent access, medication availability, or continuity of care, but these mechanisms were not measured, and residence should not be interpreted causally. Higher baseline BMQ-NCD and Family APGAR scores were associated with greater odds of good adherence, supporting the potential value of psychosocial assessment at discharge. Male sex was associated with lower odds, but this coefficient is statistically fragile because only five women were in the poor-adherence group. Residual confounding, sparse-data effects, and center-specific care patterns remain plausible.

The apparent AUC of 0.879 and optimism-corrected AUC of 0.867 indicate good discrimination within the 191 complete cases. The apparent Brier score of 0.118 and optimism-corrected calibration slope of 0.923 provide complementary information on prediction error and calibration. Bootstrap correction reduces optimism when model performance is evaluated on the development sample, but it does not establish transportability. These estimates do not demonstrate superiority to other models, readiness for resource allocation, or a validated probability threshold for clinical decisions. The nomogram is a representation of the fitted equation. Independent temporal or multicenter validation, with recalibration if necessary, is required before it is used to direct patient care.

Recent computational studies add methods absent from the original review: imbalance-aware feature selection in high-dimensional cancer data⁴¹, mHealth support for people with dementia and carers⁴², feature-selection strategies for large medical databases⁴³, ADASYN-supported deep learning for stroke occurrence prediction⁴⁴, and broader machine-learning health-monitoring frameworks⁴⁵. These are useful comparators, but their diagnostic, high-dimensional, or sensor-oriented aims differ from our 191-participant four-predictor adherence model. Future external datasets should compare prespecified logistic regression with penalized and machine-learning approaches while reporting calibration and clinical utility.

Several limitations define the scope of the findings. The single-center convenience sample limits generalizability, and complete-case analysis excluded 19 of 210 enrolled participants who did not complete follow-up; although their flow was documented, attrition bias remains possible, and no outcome values were imputed. Three assessment points cannot capture shorter-term fluctuations, and the 180-day horizon does not characterize longer-term persistence or recurrent vascular outcomes. Self-report was not cross-validated against dispensing data, pill counts, or electronic monitoring. The sample also limits detailed subgroup analysis and may yield unstable coefficients for sparse categories. Strengths include nonoverlapping intervals, explicit multi-medication rules, use of predefined demographic and day-30 psychosocial predictors, and bootstrap internal validation. Future studies should link interviews with refill or electronic measures, document respondent type, extend follow-up to 12 months, and evaluate the model in independent cohorts.

Conclusion

In conclusion, this protocol defines three post-discharge assessments of self-reported medication-taking and an interpretable day-180 model restricted to four predefined clinical and psychosocial predictors. The model showed promising internal performance in the complete-case cohort but requires requires independent external validation before clinical implementation.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Analysis Script	Study Authors	JoVE_Analysis_Script.py (Uploaded as supplementary material)
Chinese Beliefs about Medicines Questionnaire-Specific (BMQ-Specific)	Original: Prof. Robert Horne. Chinese validation: Jiang S. et al. / Nie B. et al. (as cited in text)	Authorized Chinese Version. Administered via structured interview. Permission obtained from the original developer/copyright holder.
Chinese Family APGAR Index	Original: Dr. Gabriel Smilkstein. Chinese application: Smilkstein G. et al. (as cited in text)	Validated Chinese translation. Version: 5-item scale. Open access/Public domain for clinical research.
Chinese National Institutes of Health Stroke Scale (NIHSS)	Original: National Institutes of Health (NIH). Chinese validation: Lyden P.D. et al. (as cited in text)	Validated Chinese version. 11-item clinical scale (0-42). Open access.
Matplotlib	The Matplotlib development team	Version 3.10.6. Open source (https://matplotlib.org/)
NumPy	The NumPy community	Version 2.3.3. Open source (https://numpy.org/)
Python	Python Software Foundation	Version 3.13.5 (64-bit Windows environment). Open source (https://www.python.org/)
pandas	The pandas development team (NumFOCUS)	Version 2.3.3. Open source (https://pandas.pydata.org/)
SciPy	The SciPy community	Version 1.16.2. Open source (https://scipy.org/)