Source-linked AI summary
A Comprehensive Scoping Review of Bayesian Networks in Healthcare: Past, Present and Future
Evangelia Kyrimi, Scott McLachlan, Kudakwashe Dube, Mariana R. Neves, Ali Fahmi, Norman Fenton
TL;DR
Healthcare BNs have attracted substantial research, but the field lacked a comprehensive review to organize contributions and identify gaps. This scoping review assessed decision support, modelling, and clinical usefulness across 123 papers, finding restricted use of BN reasoning and a gap between model accuracy and clinical usefulness.
Problem
The healthcare BN literature lacked a comprehensive review to organize contributions and identify research gaps, while evidence of clinical adoption remained limited.
Source: §1 · p. 2Method
The scoping review assessed decision support, BN development, and model usefulness across 123 papers proposing healthcare BNs.
Source: §1 · p. 2, §10 · p. 25Results
Most published BNs supported simple evidence-based decisions, and the review found a gap between accurate models and demonstrated clinical usefulness.
Source: §7.1 · p. 21, §10 · p. 25Takeaways & Limitations
The authors propose future research directions and recommendations to improve modelling reproducibility and BN usefulness, aiming to address the gap between research interest and clinical adoption.
Source: §10 · p. 26The selected papers may not represent the entire healthcare BN literature because relevant studies could have omitted the search keywords from their abstracts.
Source: §7.2 · p. 22
Abstract
from arXiv · showhide
No comprehensive review of Bayesian networks (BNs) in healthcare has been published in the past, making it difficult to organize the research contributions in the present and identify challenges and neglected areas that need to be addressed in the future. This unique and novel scoping review of BNs in healthcare provides an analytical framework for comprehensively characterizing the domain and its current state. The review shows that: (1) BNs in healthcare are not used to their full potential; (2) a generic BN development process is lacking; (3) limitations exists in the way BNs in healthcare are presented in the literature, which impacts understanding, consensus towards systematic methodologies, practice and adoption of BNs; and (4) a gap exists between having an accurate BN and a useful BN that impacts clinical practice. This review empowers researchers and clinicians with an analytical framework and findings that will enable understanding of the need to address the problems of restricted aims of BNs, ad hoc BN development methods, and the lack of BN adoption in practice. To map the way forward, the paper proposes future research directions and makes recommendations regarding BN development methods and adoption in practice.
1 Introduction
Although healthcare has a large and growing BN literature, no systematic or scoping review had organized the domain, and evidence of clinical adoption remained limited. This scoping review evaluates decision support, modelling approaches, and usefulness in practice.
1 Introduction
- Efficient inference algorithms and greater computational power enabled BN development for real-world decision-support problems, renewing research interest.The authors note that earlier attempts to apply Bayesian analysis to medical problems were generally considered unsuccessful because inference was computationally intractable.Source: §1 · p. 2
- The literature lacked a systematic or scoping review, despite extensive research interest and limited evidence that published BNs were adopted in practice.The authors identify a gap between a model's claimed accuracy and demonstrated clinical usefulness or impact on healthcare decisions.Source: §1 · p. 2
- The review evaluates healthcare BNs by their targeted decision support, modelling approach, and usefulness in practice.The authors describe it as the first scoping review of BNs in healthcare, to their knowledge.Source: §1 · p. 2
2 Background
Bayesian networks represent probabilistic relationships graphically and support reasoning under uncertainty. Their reasoning modes and ability to combine evidence and expert judgment motivate their use in healthcare decision support.
2.1 Bayesian fundamentals
- Bayesian inference updates prior beliefs with evidence to obtain posterior probabilities, as clinicians do when interpreting diagnostic tests and examination findings.The passage describes Bayes’ theorem as calculating a hypothesis’s posterior probability from its prior probability and the evidence likelihood under the hypothesis and its negation.Source: §2.1 · p. 2
- A BN is a directed acyclic graph whose nodes represent variables and whose arcs represent causal or influential relationships; node probability tables encode conditional distributions.Structure and parameters may be learned from data, elicited from experts and literature, or developed using both approaches.Source: §2.1 · p. 3
2.2 Reasoning process
- Observational reasoning updates unobserved variables after evidence is entered; causal reasoning follows arc direction, while diagnostic reasoning runs against it.The examples show smoking raising the probability of lung cancer from 6.4% to 10%, and a positive X-ray increasing the probability of lung cancer.Source: §2.2 · p. 3, Figure 1 · p. 4
- Interventions model hypothetical actions by fixing a variable and removing incoming edges, while counterfactual reasoning combines evidence with interventions in twin networks.Interventional reasoning requires causal relationships and does not support diagnostic reasoning from the intervened variable.Source: §2.2 · p. 4
2.3 Significance in healthcare
- BNs support healthcare decisions by modelling complex uncertain dependencies, combining data with expert judgment, and enabling interpretable diagnostic, prognostic, and intervention reasoning.Clinical decision making involves gathering evidence and making diagnostic and treatment decisions under uncertainty.Source: §2.3 · pp. 4–5
3 Motivation
The preliminary review identified substantial BN research alongside unclear development practices and negligible clinical adoption. These concerns motivated a larger review to identify issues and propose a way forward.
3 Motivation
- The healthcare BN literature was large and growing, but had seen no systematic or scoping review to summarize its contributions and research gaps.These conditions motivated undertaking a larger review.Source: §3 · p. 5
- Published BN development processes were rarely described clearly enough to be repeatable, and clinical adoption of published models was negligible.Source: §3 · p. 5
- The authors argue that absent systematic reviews may have encouraged duplicated effort and repeated methodological mistakes while leaving important gaps hidden.Source: §3 · p. 5
4 Methodology
The review searched healthcare and informatics literature, then used an analytical framework to assess decision support, modelling approaches, and model usefulness. A structured review process was intended to support consistent assessment.
4.1 Literature search
- The search covered major health and health-informatics databases and screened abstracts, publication dates, language, healthcare relevance, and decision-support or prediction focus.The broad query used Bayesian-network and probabilistic-graphical-model terms alongside medical or clinical terms; eligible papers were published from 2012–2018 in English.Source: §4.1 · p. 5
4.2 Analytical framework for characterizing BNs in healthcare
- The analytical framework identifies six objectives, while this paper reports on three: decision support, modelling approach, and model usefulness.The framework is presented as a way to characterize the healthcare BN domain and guide the review.Source: §4.2 · pp. 5–6, Figure 2 · p. 7
- The modelling assessment records BN structure, parameters, their sources, development repeatability, elicitation or learning processes, and software tools.The usefulness objective examines potential clinical usefulness against properties the authors consider necessary for a useful model.Source: §4.2 · p. 6
- Attributes were refined with decision-science input, and reviewers used a questionnaire, manual, training, and paired assessments with consensus resolution of disagreements.Two senior reviewers collaboratively resolved differing responses.Source: §4.2 · p. 7
5 Results
The review screens healthcare BN literature across decision support, modeling approaches, and usefulness in practice. Its findings describe diverse modeling choices alongside gaps in reporting and evidence of clinical use.
5.2.1 Decision support
- Observational reasoning dominated; diagnosis was the most frequent decision, while intervention was less common and actual counterfactual application appeared in one paper.Many papers learned associations or causal relationships without performing a reasoning process, which the review classifies closer to model development than model use.Source: §5.2.1 · p. 9, Figure 5 · p. 10
5.2.2 Modelling approach
- The BN development process was omitted or only implied in 48% of papers, and fewer than a quarter described or illustrated it in detail.A complete BN was provided in 53% of papers, but variables and their states were described in only 28%; collaborative papers by decision scientists and clinicians tended to be more repeatable.Source: §5.2.2 · pp. 13–14, Figure 9 · p. 15
- In 35% of papers, software was unspecified or authors used ad-hoc software; GeNIe and R were the most often used tools.Figure 10 identifies software use across the reviewed papers.Source: §5.2.2 · p. 15, Figure 10 · p. 15
5.2.3 Model usefulness
- Clinical benefits were often described, but only 5% of papers identified generalizability, fewer than 1% proposed an interface, and none showed actual clinical impact or current use.Only one paper evaluated a proposed interface; potential impact was investigated in three papers, and 34% of published BNs appeared credible under the review’s definition.Source: §5.2.3 · p. 16, Figure 12 · p. 17
6 Major Findings
The review classifies findings across decision support, BN development, and model usefulness, using three taxonomies to characterize the healthcare BN literature. It finds observational reasoning most common, development approaches often bespoke, and adoptability neglected.
6 Major Findings
- The three bidirectional taxonomies classify concepts across the review’s objectives, with connecting-line weight representing the percentage of papers associated with each relationship.They cover decision support, modeling approach, and model usefulness.Source: §6 · p. 17, Figure 16 · p. 17
6.1 Decision support
- Observational reasoning dominates the reviewed BNs, while interventional and counterfactual reasoning remain significantly underexplored.Diagnosis is the most frequent clinical decision; treatment recommendations are most frequent for interventional reasoning, and learning causal attributes is the most frequent decision context.Source: §6.1 · p. 17, Figure 17 · p. 18
6.2 Modeling approach
- More than half of published BNs are static, data-driven methods are more common than knowledge-driven ones, and hybrid-driven papers offer more generic development approaches.The review distinguishes variables, arcs, and parameters, and identifies decision scientists and clinicians as primary developer actors.Source: §6.2 · p. 18, Figure 18 · p. 19
6.3 Model usefulness
- Adoptability is the most neglected usefulness category, reflecting the scarcity of papers claiming that a BN has been adopted into clinical practice.The review groups usefulness elements under applicability, validity, and adoptability; benefit, accuracy, and credibility are most frequently observed.Source: §6.3 · p. 19, Figure 19 · p. 20