Bayesian Network Decision-Support for Severe Lower Limb Trauma
Publisher
Metadata
Show full item recordAbstract
Severe lower limb injuries are potentially devastating and pose some of the most difficult
decisions in trauma surgery. The goal is to ensure survival and reconstruct the most
functional limb possible. Ideally this is achieved by salvaging the injured limb. However,
in certain situations amputation is the safest and most effective method of achieving an
optimal outcome. Errors in these decisions may have profound consequences, yet they are
frequently based on incomplete information and uncertain risks. Furthermore, most
surgeons have limited experience making these decisions, and existing decision-support
tools are unhelpful.
The aim of this thesis was to improve the understanding of decision-making following
severe lower limb trauma, and develop accurate prognostic models that can help identify
those patients whose limb can be safely and effectively salvaged, and also identify those
for whom attempts at limb salvage would be dangerous or fail.
The rationale for amputation decisions was analysed in a cohort of severe lower limb
injuries (n = 579). Two prognostic models were designed to support difficult aspects of
these decisions. Both models were developed using Bayesian networks that combine
existing knowledge with individual patient data. The first provides early and accurate
identification (AUROC = 0.927) of patients at risk of Trauma-Induced Coagulopathy, the
principal indication for damage-control intervention. The model’s performance in new
patients, and ability to handle missing predictor information, was prospectively validated.
The second model accurately predicts the likely outcome, in terms of viability, of
attempted limb salvage. This model outperformed the most widely used decision-support
tool, the Mangled Extremity Severity Score (AUROC 0.932 versus 0.723; P < 0.0001).
These Bayesian network tools accurately quantify critical risks that make rational
judgement on the safety and effectiveness of interventions possible. This information
enables individualised and evidence-based decisions, at a time when decision-making is
most effective.
Authors
Perkins, Zane BrendanCollections
- Theses [3822]