Machine Learning Supports Automated Digital Image Scoring of Stool Consistency in Diapers

Area of Science:

• Pediatric gastroenterology
• Medical imaging analysis
• Machine learning applications in healthcare

Background:

• Accurate stool consistency assessment is crucial for monitoring infant health but challenging in non-toilet-trained children.
• Current methods often rely on subjective caregiver reporting, leading to inconsistencies.
• Automated classification of stool consistency from diaper images presents a potential solution.

Purpose of the Study:

• To evaluate the feasibility of using machine learning (ML) for automated stool consistency classification from smartphone photos of diapers.
• To develop and test a proof-of-concept ML model for objective stool scoring.

Main Methods:

• Collected 2687 diaper photos from 96 infants (<24 months).
• Stool consistency was scored by participants, researchers, and a healthcare professional using the Brussels Infant and Toddler Stool Scale.
• A deep convolutional neural network model was trained using transfer learning on a subset of the photos.

Main Results:

• Inter-rater agreement among human scorers varied (researchers: 77.5%).
• The ML model achieved 60.3% exact agreement with the final expert score on test photos.
• Agreement improved to 77.0% when using a 4-class grouping of the stool scale.

Conclusions:

• Automated stool consistency scoring via ML from diaper photos demonstrates robust agreement with human assessments.
• This ML-driven approach overcomes limitations of subjective caregiver reporting.
• The framework has potential applications in clinical trials and home-based infant health monitoring.