A proof-of-concept study to construct Bayesian network decision models for supporting the categorization of sudden unexpected infant death

Abstract

Sudden infant death syndrome (SIDS) remains a leading cause of infant death in high-income countries. Supporting models for categorization of sudden unexpected infant death into SIDS/non-SIDS could reduce mortality. Therefore, we aimed to develop such a tool utilizing forensic data, but the reduced number of SIDS cases renders this task inherently difficult. To overcome this, we constructed Bayesian network models according to diagnoses performed by expert pathologists and created conditional probability tables in a proof-of-concept study. In the diagnostic support model, the data of 64 sudden unexpected infant death cases was employed as the training dataset, and 16 known-risk factors, including age at death and co-sleeping, were added. In the validation study, which included 8 new cases, the models reproduced experts' diagnoses in 4 or 5 of the 6 SIDS cases. Next, to confirm the effectiveness of this approach for onset prediction, the data from 41 SIDS cases was employed. The model predicted that the risk of SIDS in 0- to 2-month-old infants exposed to passive smoking and co-sleeping is eightfold higher than that in the general infant population, which is comparable with previously published findings. The Bayesian approach could be a promising tool for constructing SIDS prevention models.

Figure 1

Ratios of sudden unexpected infant death (SUID) subcategories to total SUID cases in the 47 prefectures in Japan from 2012 to 2018. SUID was defined as a set of the following three subcategories: SIDS, accidental asphyxia, and unknown causes of death. SIDS, R95 (SIDS) in ICD-10; accidental asphyxia, the combination of W75 (accidental suffocation and strangulation in bed), W78 (inhalation of gastric contents), and W79 (inhalation and ingestion of food, causing obstruction of respiratory tract); unknown causes of death, the combination of R96 (other sudden death, cause unknown), R98 (unattended death), and R99 (other ill-defined and unspecified causes of mortality). SUID, sudden unexpected infant death; ICD-10, International Classification of Diseases, 10th Revision; SIDS, sudden infant death syndrome.

Figure 2

Flow chart of case enrollment and exclusion.

Figure 3

A Bayesian diagnostic support model for SIDS. This model reflects the relationship among risk factors leading to death (cf. Supplementary Fig. S2). A conditional probability table was created for each factor. Including the presence or absence of each factor enables the calculation of SIDS diagnosis probability. SIDS, sudden infant death syndrome.

Figure 4

A Bayesian onset-predictive support model for SIDS. A conditional probability table was incorporated for each factor. Including the presence or absence of each factor produces a SIDS-onset probability as an annual incidence rate per 1000 of the population. The prior probability of annual SIDS incidence in the general population was 0.3/1000 live births calculated from vital statistics of a population survey^, and our 64 cases. SIDS, sudden infant death syndrome.