gms | German Medical Science

Background: Beta-blockers are cornerstones in the secondary prevention after myocardial infarction and in the treatment of (stable) systolic heart failure (i.e., heart failure with reduced ejection fraction). In both conditions, readmissions after an index hospitalization are manifestations of progressive disease that could potentially be reduced or avoided with appropriate medication. The particular beta-blocker and its dosage are often chosen subjectively according to concomitant diseases or pragmatically according to tolerability. This is not surprising, because there is an objective lack of clear systematic evidence comparing purely beta1-selective agents such as metoprolol and bisoprolol or the additional alpha-blocking and vasodilator carvedilol. While group comparisons show ambiguous results, the question arises whether the response to beta-blockers differs between subtypes. In posthospitalization situations for myocardial infarction (A) and heart failure (B), we pursued the working hypothesis that heterogeneous treatment effects (HTE) exist in the response to metoprolol (1), bisoprolol (2), and carvedilol (3) and that an individualized treatment recommendation can be derived from an individual response to substance and dose.

Materials and Methods: Health insurance claims data from AOK Baden-Württemberg from 2011 to 2016 were processed, so that specific rehospitalizations (also considering symptoms, manifestation of index disease, complications, and potential adverse drug reactions) were determined in the follow-up period of patients after an index stay for myocardial infarction or heart failure [1]. In a new-user design, we selected patients newly initiating treatment with one of the aforementioned beta-blockers (1-3) after their index stay (patients treated with other substances and or switching between substances were excluded from the analysis set). Starting from the year before the index hospitalization, drug exposure was determined in a time-dependent manner by first determining the mean approximated dose of each patient from the prescription patterns of dosage units and package sizes and then by using the approximated dose to calculate the prescription durations [2]. When appropriate, half of the population mean days of supply was added as the grace period. Thus, for the myocardial infarction subset (A), exposure windows were obtained considering the last episode before an event or the end of the observation period according to the while-on-treatment strategy (Figure 1 [Fig. 1]). For the heart failure subset (B), the dose of the last episode was extracted (target dose). In statistical models, the readmission time or end of observation (censoring) was considered as the dependent variable; we substituted the observed readmission times by jackknife pseudo-observations, which were then used as a quantitative response variable for predictive modeling of HTE [3]. From the domains of demographics, medical history, comedication, comorbidities, and multi-nominal propensity scores for beta-blocker prescriptions, 44 (A) and 28 (B) effect modulators were considered. Time-dependent recommendations were calculated using Bayesian additive regression trees (BART package in R version 4.0.5), whereas individual dose-response relationships were estimated using generative adversarial networks (SCIGAN framework in Python version 3.8.5).

Results: After an index hospitalization for myocardial infarction (A), 3305 patients (mean age: 78.3 years) were newly treated with a beta-blocker (57.1% metoprolol, 38.9% bisoprolol, 4.0% carvedilol). After an index hospitalization for heart failure (B), 5023 patients (mean age: 81.6 years) were newly treated with a beta-blocker (52.9% metoprolol, 38.9% bisoprolol, 7.2% carvedilol). After the index stay for myocardial infarction, the episode before event or end of observation period showed that guideline-based treatments also increased prominently in comedication (Figure 2A [Fig. 2]). Among the ultimately derived target doses were typical dose levels (after standardization using defined daily doses, DDDs), and only few patients were treated above one DDD of the respective agent (Figure 2B [Fig. 2]). Within 365 days, 2017 readmissions for myocardial infarction and 498 readmissions for heart failure were observed. Considering individual risks and probabilities of success, clear patterns in individualized treatment recommendations emerged from HTE modeling depending on the time for readmission: Concerning readmission for myocardial infarction (A), up to 42.5% of patients would receive a model-based recommendation for carvedilol, which was stable over time (Figure 3A [Fig. 3]). Among beta1-selective agents, metoprolol appeared to gain importance with increasing time after index hospitalization. In heart failure (B), plausible individual dose-response relationships could be determined by artificial intelligence and causal inference methods (Figure 3B [Fig. 3]). While a similar pattern emerged in the mean curves, inter-individual variability was much more pronounced in the individual carvedilol curves.

Conclusion: Inter-individual variability in response to beta-blockers of different subtypes could potentially guide treatment choices in the future. If such HTE could be determined automatically from routine data by model-based recommendations, this would be a useful tool for decision support. The present work suggests that a much larger proportion of patients after myocardial infarction and with heart failure could be treated with carvedilol and, given that treatment is provided to the most susceptible patients, this could also reduce clinical events such as readmissions. Using routine data for such individualized recommendations closes the loop of medical knowledge, where each individual case in everyday use can also provide new evidence for decisions in future patients.