Bayesian networks are graphical models proved to be effective in medical domains where diagnostic and prognostic inference is required. However, medical problems involve a number of variables representing clinical judgements, which are never observed within an ordinary clinical setting. Hence, the elicitation of conditional distributions is typically demanding due to the huge number of parameters to be considered and to partial expert knowledge. Bayesian parameter estimation based on collected data may be inaccurate, as latent variables often entail partial model identification. Informative prior distributions are helpful in obtaining an effective working model. In this paper, we propose a new elicitation framework involving quantitative latent variables in which parsimony is a key issue but enough expressiveness is preserved. Our approach is specifically focused on medical domain knowledge, where informative prior distributions are elicited by asking expert physicians to state their beliefs about features actually retrievable in medical practice.
[1]
Luca Antiga,et al.
Bayes pulmonary embolism assisted diagnosis: a new expert system for clinical use
,
2007,
Emergency Medicine Journal.
[2]
Steen Andreassen,et al.
MUNIN - A Causal Probabilistic Network for Interpretation of Electromyographic Findings
,
1987,
IJCAI.
[3]
Davide Luciani,et al.
Automated interviews on clinical case reports to elicit directed acyclic graphs
,
2012,
Artif. Intell. Medicine.
[4]
D. Heckerman,et al.
Toward Normative Expert Systems: Part I The Pathfinder Project
,
1992,
Methods of Information in Medicine.
[5]
D. Heckerman,et al.
,81. Introduction
,
2022
.