Artificial intelligence-based non-invasive bilirubin prediction for neonatal jaundice using 1D convolutional neural network

Abstract

Neonatal jaundice, characterized by elevated bilirubin levels causing yellow discoloration of the skin and eyes in newborns, is a critical condition requiring accurate and timely diagnosis. This study proposes a novel approach using 1D Convolutional Neural Networks (1DCNN) for estimating bilirubin levels from RGB, HSV, LAB, and YCbCr color channels extracted from infant images. Initially, each color channel is treated as a time series input to a 1DCNN model, facilitating bilirubin level prediction through regression analysis. Subsequently, RGB feature maps are combined with those derived from HSV, LAB, and YCbCr channels to enhance prediction performance. The effectiveness of these methods is evaluated based on Root Mean Squared Error (RMSE), R-squared (R²), and Mean Absolute Error (MAE). Additionally, the best-performing model is adapted for classification of jaundice status. The results show that the integration of RGB and HSV color spaces yields the best performance, with an RMSE of 1.13 and an R² score of 0.91. Moreover, the model achieved an impressive accuracy of 96.87% in classifying jaundice status into three categories. This study provides a promising non-invasive alternative for neonatal jaundice detection, potentially improving early diagnosis and management in clinical settings.

Keywords: Bilirubin level prediction; Image processing; Neonatal jaundice; One dimensional convolutional neural network.

Fig. 1

Diagram outlining the stepwise approach of the proposed framework.

Fig. 2

Left: Neonate, 5 days old, with bilirubin 14.6 mg/dL (abnormal). Right: Neonate, 5 days old, with bilirubin 3.9 mg/dL (normal).

Fig. 3

How to integrate the obtained features with the CNN model. For example, two color spaces RGB and HSV are given.

Fig. 4

The values of each pixel in 4 color spaces show the images on the left of a 4-day-old normal baby with a bilirubin of 7 and on the right of a 3-day-old hyperbilirubin baby with a TSB of 17.1.

Fig. 5

Confusion matrix of best classification result across three classed.

Fig. 6

Comparison of performance metrics (Accuracy, Precision, Recall, and F1 Score) for bilirubin level classification across different color spaces.