Evaluation of a Bayesian decision network for diagnosing pyloric stenosis

Area of Science:

• Pediatric diagnostics
• Medical decision support systems
• Bayesian networks in medicine

Background:

• Clinical assessment for pyloric stenosis often lacks accuracy, leading to unnecessary ultrasounds.
• Pyloric stenosis diagnosis in infants frequently relies on clinical evaluation, which has limitations.

Purpose of the Study:

• To assess the feasibility of a Bayesian network for enhancing pyloric stenosis diagnosis accuracy.
• To determine if a Bayesian decision network can improve diagnostic precision for infantile pyloric stenosis.

Main Methods:

• A Bayesian decision network was trained using data from 88 infants and tested on 28 infants with suspected pyloric stenosis.
• Physicians predicted pyloric stenosis probability with and without the Bayesian network, with accuracy measured by ROC curves and H-L statistics.

Main Results:

• Physicians using the Bayesian network achieved higher diagnostic accuracy (ROC 0.973) compared to those without (ROC 0.882).
• The Bayesian network approach led to a 22% reduction in ultrasounds ordered, without missing any pyloric stenosis cases.
• Calibration was significantly better with the network (H-L c-statistic 3.9) versus without (H-L c-statistic 24.3).

Conclusions:

• Bayesian decision networks show promise in improving diagnostic accuracy for pyloric stenosis in infants.
• Implementing this decision tool may allow for a safe reduction in imaging utilization for suspected pyloric stenosis.