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

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

Stool consistency or changes in stool consistency can be a common source of parental anxiety, prompting them to seek help of their health care professionals.
In daily pediatric practice, stool consistency scoring is important for assessing gastrointestinal health and disorders.

The digital innovation team at our Precision Nutrition D-Lab have developed a machine learning (ML) based algorithm that automatically assesses stool consistencies of non-toilet trained children from smartphone photos taken by parents. Until now, smartphone photos have not been well studied for potential use in this type of medical image recognition.

The feasibility of using machine learning to classify stool consistencies from diaper images was investigated in the study ‘Machine Learning Supports Automated Digital Image Scoring of Stool Consistency in Diapers’, published in the peer-reviewed Journal of Pediatric Gastroenterology and Nutrition.

Dr. med. Thomas Ludwig, Principal Scientist and 1^st author of the study said “The use of smartphone photos is less well explored in medical image recognition. There are a lot of smartphones and diapers, and we found it logical to explore the combination of both. As a result we conducted a study where, for the first time, deep learning was applied to automatically assess stool consistencies of non-toilet trained children from smartphone photos taken by mothers. More than 90% of mothers that participated in the study concluded that an automated tool to score stool consistencies would be helpful. This study also illustrates opportunities for using smartphones to rapidly generate and annotate photo datasets relevant for clinical use.”

In the study, a database of 2687 smartphone photos of used diapers from 96 children younger than 24 months was collected. The stool consistency of each of the photos was assessed by researchers. Then the image database was randomly split. 2478 photos were used to train the machine learning algorithm. 209 photos not “seen” by the algorithm were used to test its accuracy versus the stool consistency assigned by the researchers.

Stool quality from each photo was assessed using the ranking of 7 types of stool consistencies of the Brussels Infant and Toddler Stool Scale (BITSS).
To ensure flexibility towards adaptations of the BITTS scale or other scales, the ML model was developed on the initial order of 7 BITSS types.
The validated BITSS, however, clustered some of these initial 7 types of stool consistencies into 4 groups.
The ML model showed a 60,3% accuracy between model-based and researcher classification when the 7-class grouping was applied.
Performance of the higher level 4-class grouping of the 7 BITSS types, showed a 77% agreement of the proof of concept machine learning algorithm and the researchers.

This study shows the strong potential of using digital innovation, such as combining a smartphone app with machine learning, to quickly and reliably generate and annotate photo datasets in clinical use. In addition is has potential for use in clinical studies and home assessments.

Agathe Foussat, Head of the Digital Innovation Program at the Precision Nutrition D-lab explains: “Digital tools, like the ‘Stool Tracker’ algorithm, can offer parents another valuable source of information on their baby’s or toddler’s gut health, when parents can at times struggle to book in person, face-to-face appointments with their healthcare professional.”

Read more about the study here: Machine Learning Supports Automated Digital Image Scoring of… : Journal of Pediatric Gastroenterology and Nutrition