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

The aim of the study was to develop a deep convolutional neural networks (CNNs) algorithm for automated assessment of stool consistency from diaper photographs and test its performance under real-world conditions.

Diaper photographs were enrolled via a mobile phone application. The stool consistency was assessed independently according to the Brussels Infant and Toddler Stool Scale (BITSS) by paediatricians. These images were randomised into a training data set and a test data set. After training and testing, the new algorithm was used under real-world conditions by parents.

There was an overall agreement of 92.9% between paediatricians and the CNN-generated algorithm. Post hoc classification into the validated 4 categories of the BITSS yielded an agreement of 95.4%. Spearman correlation analysis across the ranking of 7 BITSS photographs and validated 4 categories showed a significant correlation of rho = 0.93 (95% CI, 0.92, 0.94; p < 0.001) and rho = 0.92 (95% CI, 0.90, 0.93; p < 0.001), respectively. The real-world application yielded further insights into changes in stool consistency between age categories and mode of feeding.

The new CNN-based algorithm is able to reliably identify stool consistency from diaper photographs and may support the communication between parents and paediatricians.