Generation and application of a convolutional neural networks algorithm in evaluating stool consistency in diapers

Area of Science:

• Medical Artificial Intelligence
• Pediatric Gastroenterology
• Digital Health

Background:

• Assessing infant stool consistency is crucial for monitoring health.
• Current methods rely on subjective visual assessment by parents and clinicians.
• Objective and standardized tools are needed for accurate evaluation.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) algorithm for automated stool consistency assessment using diaper photographs.
• To evaluate the algorithm's performance in real-world conditions.

Main Methods:

• Diaper photographs were collected via a mobile app.
• Stool consistency was independently assessed by pediatricians using the Brussels Infant and Toddler Stool Scale (BITSS).
• A CNN algorithm was trained on a data set of images and subsequently tested.

Main Results:

• The CNN algorithm achieved 92.9% overall agreement with pediatricians' assessments.
• Classification into 4 validated BITSS categories showed 95.4% agreement.
• Strong correlations (rho=0.93 and 0.92) were observed between the algorithm and BITSS rankings.

Conclusions:

• A CNN-based algorithm reliably determines stool consistency from diaper images.
• This technology can potentially enhance communication between parents and healthcare providers.
• The tool offers insights into stool consistency variations related to infant age and feeding.