An In-Depth Review Of The Uses Of Bayesian Methods In Biostatistics In Drug Safety And Pharmacovigilance.

Bayesian stats are getting this really hip and versatile to combine what we already have with new data in the context of a probabilistic logic feel in the modern biostatistics.^[13,16] In contrast to the bare and simple frequentist model which simply examines the data and follows fixed numbers, the Bayesian model considers parameters to be random variables that get or refined with more evidence. It is a living model and it keeps updating us and provides us with a clear picture whether we are sure about something.^[1]

Considering the past couple of decades Bayesian methods have ceased to be roughly a theory class discussion and now are real-life clinical trial design, drug safety test, pharmacovigilance, dose-response orientation and meta-analysis^[2,21]. They are increasingly becoming accepted in clinical and regulatory decision-making, even being backed by the most reputable authorities, such as FDA and EMA, particularly when dealing with small sample sizes, that is, early-phase studies or rare-disease trials. ^[14,32]Conventional frequentist methods could not want to somewhat suffocate on low-frequency events or small samples, which is the typical case with drug safety and post-marketing surveillance. Using Bayesian methods solves that, severely estimating and making sound inferences even exiguous data sets, by overlaying prior distributions to add new data, making the predictions more robust.Bayesian hierarchical models in fact nail the accuracy of adverse drug reaction (ADR) detection in the context of pharmacovigilance. They allow borrowing of strength both among drugs, between populations and among datasets, such as WHO VigiBase and the FDA FAERS, to enhance signal detection and reduce false positives. It is similar to a larger team. ^[6,8]

Adaptive clinical trial designs were also transformed using the Bayesian techniques. Trials are more ethical, more efficient, and more informative with real-time corrections (as with interim data) to sample size, dose plans, randomization ratios, etc. and particularly helpful in oncology, vaccine development and rare diseases.

Bayesian inference is theoretically based on the Bayes’ theorem ^[11,24], which was originally written out during the 18th century by Reverend Thomas Bayes. It mathematically expresses the use of prior knowledge in the reinforcing of new evidence:

P(θ?D)=P(D?θ) P(θ)P(D)

where:

- θ is the parameter,

- D is the data,

- P ( 0 ) is the posterior probability,

- P(D | θ) is the likelihood,

- P(θ) is the prior distribution.

The denominator P (D) normalizes the posterior such that it adds up to one. This is the essence of Bayesian thinking, which allows scientists to take a mixture of knowledge, information, or facts gained throughout history and apply them to dynamic analysis.

More recent computational advances have rendered Bayesian methods quite usable. Algorithms such as MCMC, HMC and VI allow us to fit complex hierarchical models which in the past were computational nightmares.^[10,20] The Bayesian modeling has been made easier through software like Stan, JAGS, BUGS and PyMC3/4, which have provided approachable probabilistic programming environments. ^[26–28]

Bayesian approaches tend to perform better than standard disproportionality tests (e.g. Reporting Odds Ratio, Proportional Reporting Ratio) in pharmacovigilance when the data is sparse. These problems are addressed in models such as Empirical Bayes Geometric Mean (EBGM) and Bayesian Confidence Propagation Neural Network (BCPNN) which incorporate prior information, thus resulting in improved and more confident signal detection.^[6] In addition, Bayesian networks and dynamic models are used in measuring the multi-level uncertainty and causality in drug-event associations.

In addition to inference, Bayesian predictive model and decision theory have reverted the way clinical and regulatory decisions are made. The Bayesian decision theory provides a consistent scheme applied in the estimation of the expected utilities of various decisions made in the face of uncertainty which is essential to benefit-risk analysis, the Bayesian models are structured to take both efficacy and safety data to aid in evidence-based regulatory appraisals. Bayesian predictive checks also enable us to model a possible trial behaviour and check the soundness of the model on a pre-study or within study basis.

With the biomedical research field moving towards real world data, genomics and personalized medicine, the Bayesian methodological approach can synthesize heterogeneous information easily, and can incorporate missing data, latent variables, and hierarchies. Bayesian biostatistics is, technically, not a new tool, but a change of philosophy: uncertain, integrative, driven.

This review intends to recapitulate the current advances in the Bayesian uses within the biostatic context, particularly in terms of drug safety testing, pharmacovigilance, and regulatory science with the aim of highlighting implementation methodologies, computer programs, and feasible issues to wider application.

METHODS:

Search methodology

The search methodology of the literature was rather simple because it was based on the key words provided by the authors of the sources used.

We searched peer-reviewed research papers, reviews, reports which discuss Bayesian techniques in biostatistics, pharmacovigilance, and drug safety. Search was conducted in accordance with the systematic review rules to make it rigorous and have only high-quality evidence.

We also searched the references of the related reviews to identify the additional sources not picked by databases. Methodological and applied literature was cherry-picked to present the extensive application of Bayesian methods to biomedical and regulatory issues.

Inclusion and Exclusion Criteria

• Inclusion Criteria:
  • English peer-reviewed, review, or meta-analysis articles.
  • In vivo descriptions of the application, development, or assessment of Bayesian methods in drug safety, analysis of efficacy, pharmacovigilance, or biostatistics.^[17,21]
  • Conferences that mention case studies, computational methods (such as MCMC, hierarchical methods), or regulatory applications with the FDA, the EMA or WHO-UMC.
• Exclusion Criteria:
  • Not in English or peer-reviewed articles (e.g., preprints).
  • Theoretical work or mathematical work which is not directly related to biomedical science.^[11]
  • Those articles that appear not to have a more in-depth analysis of Bayesian analysis.
  • Research not related to data analysis or not in the health sciences.

Process of screening and selection

Our search began with the identification of studies and we identified 1, 247 papers. We eliminated duplicates after which we were left with 982 distinct records. The number was reduced to 317 potentially useful papers when we were filtering titles and abstracts to arrive at the relevant ones. Then, having checked the entire texts in accordance with our inclusion criteria, we selected 68 studies to dive into them.

Figure 1 Selection process

We recapped the entire selection process in a PRISMA based flow diagram (Figure 1).^[18] The papers were classified according to their general field of application:

• Bayesian statistics in clinical trials,
• Bayesian drug safety
• Pharmacovigilance model
• Bayesian dose response or survival.

Data Extraction and Analysis

The data extraction and analysis is viewed as a data mining since it is operated automatically to determine which data is considered important and that which is not in the study problem.

In each of the chosen studies, we extracted the main information, such as:

• Authors, date of publication and location of study and research;
• Bayesian approach type (e.g. hierarchical model, MCMC, Bayesian network);
• Application area (clinical trials, risk benefit assessment, pharmacovigilance, etc.);
• Significant results, software applied, and the benefits or limitations mentioned.

Due to the high diversity of studies we have used, some of them have simulations, some develop an algorithm, some describe a real-life case study, we have chosen to conduct a narrative descriptive synthesis, rather than a quantitative meta-analysis.^[34] Concentration remained on how Bayesian approaches enhance the quality of inference, model transparency and making decisions in biostatistics practice.

Quality Assessment

In order to ensure the credibility of the evidence, we measured the quality of the methodology of each conclusion by applying the criteria described in Spiegelhalter et al. (2004)^[4] and Gelman et al. (2013).^[1] We looked at:

• Transparency of model assumptions;
• Vindication and suitability of precedents;
• Adequacy of model checking (e.g., posterior predictive diagnostics);
• The reproducibility of computations.

We focused on papers that contained sensitivity analysis or MCMC convergence diagnostic or explicit model validation.

Ethical Considerations

We used all the data that were found in publicly available and peer-reviewed sources. It did not directly involve any human or animal subjects. In achieving the review, we went by all the rules and ethics without engaging in plagiarism, fabrication of data and also unnecessary publication.

RESULTS AND DISCUSSION:

Foundation of Bayesian Biostatistics

Using Bayesian statistics, we can make one coherent inference because probability is regarded as the degree of belief, or degree of confidence, even though this is the frequency level in its measurement. In contrast to the frequentist perspective, according to which probability is a long run relative frequency of events, the Bayesian model allows us to state our level of belief about an unknown parameter based on the evidence that we are presently in possession of. This distinction comes in painfully handy in biostatistics, and in practice in biostatistics when we frequently have uncertainty, inadequate data, and ethical considerations restricting experimentation.

The Bayesian inference lies at the heart of the Bayesian theory, and it inform us of how to revise our prior beliefs regarding a parameter when we observe new data:

P(θ?D)=P(D?θ) P(θ)P(D)

In this case, the left term is the posterior distribution, which displays what we know after updating of seeing the data; the second term is the likelihood, which is what we know or expect before seeing the new data; and the last term is the prior, which is the sum total of what we know or are expecting to see before we look at the new data. P (D) is simply a normalizing constant that ensures that probabilities sum up to one.

Since we obtain the complete posterior distribution, we are able to characterize uncertainty with regards to parameters without depending upon single point estimates. Bayesian inference implies credible intervals to us almost automatically, i.e. direct statements of probability that a parameter lies in a certain range. An example is that a 95 credible interval of treatment effect implies that with the data and the model, it has a 95 probability that the true effect is in that what is contained by the interval. This interpretation is more intuitive than the frequentist interpretation of confidence interval which simply discusses frequencies of coverage in the long-run.

The place of Priors and Posteriors

The previous distribution is a true significance in Bayesian analysis. Priors can be:

Informative, as a result of a prior study, pharmacokinetic model, or expert knowledge, or non-informative (or weakly informative) that are constructed to be influenced have a minimum impact on the posterior.

Informative priors may also be used to stabilize estimates when the sample of both variables in a study is small (that is, like in early-phase clinical trials or research on rare diseases). As an illustration, using a priori information on dose-toxicity mechanisms in dose-escalation studies may help to minimize the level of uncertainty in determining safe dose levels.

Once we observe the data, the posterior combination of the prior and the new evidence provides us with more realistic as well as robust estimates. Predictive inference is also based on the posterior, which enables us to simulate data in the future. This predictive power implies that Bayesian models can learn without full information and create the foundation of adaptive and sequential trial designs with the interim outcomes guiding the research.

Multilevel and Hierarchical modeling

Hierarchical (multilevel) modeling is considered to be one of the coolest applications of Bayesian statistics as applied to biostatistics. These models are used to deal with variation on a variety of levels: patients, clinical sites, treatment centres, or studies, and are able to borrow strength across related groups.

The parameters of each subgroup of the hierarchy are assumed to follow some hyperpriors (random variables). A two-level model will be described as a generic one and will be of the following form:

yi∼f(yi?θi),θi∼g(θi??)

where:

• yi

  

  represents the observed data for subgroup i

  

• θi

  

  denotes the parameter specific to subgroup i

  

• ?

  

  represents hyperparameters that govern the higher-level variation across subgroups

This model is particularly useful in pharmacovigilance whereby the adverse events are different drugs, population and geography. It is by aggregating data between subgroups that hierarchical models provide us with consistent approximations of drug safety signals in situations where the data is sparse at the individual-level. We have observed these methods in successful use in post-marketing surveillance systems, such as FAERS and VigiBase, to detect rare or emergent ADRs.

The computational developments are as follows

The practical implementation of Bayesian inference was previously difficult due to the difficulties in computation. Up to the 1990s, it was only possible to obtain closed-form solutions to simple conjugate models. The game has transformed with the Markov Chain Monte Carlo (MCMC) algorithms that allow us to compute accurately numerically the complex posterior distributions.

Although modern Bayesian inference is dominated by classical MCMC algorithms, such as Gibbs sampling and Metropolis Hastings. They sample the posterior distributions repeatedly, making it possible to estimate not only in case there is no closed-form solution. Hamilton Monte Carlo (HMC) and No -U-Turn Sampler (NUTS), as introduced by Stan, have much more recently achieved much faster sampling rates and convergence in high-dimensional spaces.

VIs and SMC are other emerging tools offering more rapid approximations which handle large biomedical data. Combined with open-source packages such as JAGS, BUGS, PyMC3/PyMC4, and TensorFlow Probability these developments have made Bayesian models affordable to biostatisticians, data scientists with less advanced computational skills.

Strengths in Biostatistical Uses

Bayesian procedures have several strengths in terms of the biostatistical studies or in regulatory science:

Incorporation of preknown data – We are able to import historical or professional information which is useful when we have limited samples or infrequent occurrences.

Direct Interpretation of Uncertainty Credible intervals provide us with statements of clarity about parameters in terms of probability, so you are certain of how certain you are.

Flexibility in modeling Bayesian is used when dealing with complex data structure, missing data, measurement error and hierarchical dependencies and is therefore incredibly useful when dealing with data that are not simple.

Adaptive and Sequential Learning In this case, we are able to update the model in real time since it supports adaptive clinical trial designs as well as allows us to make decisions in real time.

Decision-Theoretic Framework- The Bayesian decision theory more or less formalizes the risk benefit analysis, which is just right in regulatory expectations of transparent and evidence-based decision-making.

The following are some of the strengths of Bayesian techniques which render the techniques very useful in clinical trial monitoring, pharmacovigilance, dose finding, and post-marketing drugs safety. Bayesian inference enhances the transparency, reproducibility and strength of biomedical decision making by providing you full posterior distributions rather than only giving a single point value.

Problems and Future Projections

Although there is an increased adoption of Bayesian approaches, they have some obstacles that keep them out of full compliance regulations. Key concerns include:

Selection and justification of prior- Regulators seek transparent explanations and sensitivity analyses of the fact that the priors are not being used to skew the result.

Model validation and convergence, posterior predictive, and convergence diagnostics (such as Gelman Rubin statistics) are musts.

Computational complexity - Large hierarchical models are able to consume large amounts of computational resources even with recent developments.

Communication and interpretation - Not only the clinicians but the regulators also might require training in order to properly read Bayesian probabilities.

The future research is to design the standard reporting framework of Bayesian analysis (think CONSORT but with Bayesian studies), to increase education materials and establish open-source in columns that can reproduce and be reproducible work workflow with the goal of Bayesian work. The integration of real-world data (RWD) and machine-learning methods also appears as one of the viable frontiers of personalized and precision medicine.

Table 1