[Paper Review] Product risk assessment: a Bayesian network approach

Joshua L. Hunte, Martin Neil+1 more|arXiv (Cornell University)|Jan 1, 2020

Risk and Safety Analysis25 references2 citations

TL;DR

This paper proposes a Bayesian network (BN) model for systematic product risk assessment that overcomes key limitations of RAPEX, such as poor handling of uncertainty and lack of causal reasoning. By integrating hazard data, usage patterns, manufacturing processes, and risk perception, the BN method quantifies injury probabilities and risk tolerability even with sparse or no empirical data, demonstrating superior flexibility and interpretability in case studies on a teddy bear and an uncertified kettle.

ABSTRACT

Product risk assessment is the overall process of determining whether a product, which could be anything from a type of washing machine to a type of teddy bear, is judged safe for consumers to use. There are several methods used for product risk assessment, including RAPEX, which is the primary method used by regulators in the UK and EU. However, despite its widespread use, we identify several limitations of RAPEX including a limited approach to handling uncertainty and the inability to incorporate causal explanations for using and interpreting test data. In contrast, Bayesian Networks (BNs) are a rigorous, normative method for modelling uncertainty and causality which are already used for risk assessment in domains such as medicine and finance, as well as critical systems generally. This article proposes a BN model that provides an improved systematic method for product risk assessment that resolves the identified limitations with RAPEX. We use our proposed method to demonstrate risk assessments for a teddy bear and a new uncertified kettle for which there is no testing data and the number of product instances is unknown. We show that, while we can replicate the results of the RAPEX method, the BN approach is more powerful and flexible.

Motivation & Objective

• To address the limitations of RAPEX in handling uncertainty and incorporating causal explanations in product risk assessment.
• To develop a generic, systematic Bayesian network model that integrates hazard, usage, manufacturing, and risk perception data for improved risk estimation.
• To demonstrate the model’s capability in assessing risk for products with no testing data or unknown instance counts, such as a new uncertified kettle.
• To provide a normative, interpretable method for risk evaluation that supports evidence-based regulatory decisions.

Proposed method

• The proposed BN model uses conditional probability tables and causal structure to represent relationships between risk factors, including hazard occurrence, usage patterns, and injury severity.
• Priors are assigned to uncertain variables such as number of product uses, usage deviation rates, and product instance counts to reflect prior knowledge.
• Bayesian inference is applied to update probabilities based on evidence, including hypothetical injury reports and process data.
• The model estimates the probability of major and minor injuries per demand, hazard occurrence, and overall risk levels.
• Risk tolerability is assessed by combining injury probabilities with consumer utility and regulatory intervention thresholds.
• The model supports scenario analysis by varying inputs (e.g., number of injuries reported) to evaluate risk outcomes under different assumptions.

Experimental results

Research questions

• RQ1 How can a Bayesian network model improve the handling of uncertainty in product risk assessment compared to RAPEX?
• RQ2 In the absence of testing data, can a BN model reliably estimate product risk using prior knowledge and causal relationships?
• RQ3 How does the inclusion of usage behavior, manufacturing processes, and risk perception affect risk estimation accuracy?
• RQ4 Can the BN model support decision-making on government interventions such as product recalls with incomplete data?

Key findings

• For the uncertified kettle in Scenario 1, the BN model estimated a mean major injury probability of 0.005 and minor injury probability of 0.01, predicting 375 major and 750 minor injuries across 50,000–100,000 units.
• In Scenario 1, the model classified the risk level as 'very high' and recommended government intervention with uncertainty, despite no reported injuries.
• In Scenario 2, where one major injury was reported, the model estimated a much lower mean injury probability of 0.00004 for major injuries and recommended no intervention with low uncertainty.
• The BN model demonstrated that prior knowledge and causal modeling can yield more reliable risk assessments than RAPEX, especially when empirical data is absent or limited.
• The model successfully estimated risk for a product with no testing data, showing that BNs can handle incomplete data and revise estimates as new evidence emerges.
• The risk tolerability distribution was centered at 'low' in Scenario 1 and ranged from 'high' to 'very high' in Scenario 2, reflecting the model’s sensitivity to input assumptions and evidence.