Multimodal feature analysis for automated neonatal jaundice assessment using machine learning

Area of Science:

• Neonatal care
• Artificial Intelligence
• Medical imaging

Background:

• Neonatal jaundice monitoring is resource-intensive.
• Current AI methods use either image or clinical data, but not both.
• A systematic comparison of multimodal feature contributions is lacking.

Purpose of the Study:

• To develop an AI model combining multimodal features for accurate neonatal jaundice assessment.
• To identify an optimal feature set for jaundice prediction.
• To compare the contributions of different data modalities and features.

Main Methods:

• Collected clinical data and skin images from 633 neonates.
• Generated 460 features across 4 categories.
• Utilized tree-based models, including Light Gradient Boosting Machine (LGBM), and SHapley Additive exPlanation (SHAP) for feature importance analysis.

Main Results:

• The LGBM model with 140 features achieved an RMSE of 2.0477 mg/dL and a Pearson correlation of 0.8435.
• Combined multimodal data improved performance by over 10% compared to single modalities.
• Top 30 SHAP features reduced dimensionality while maintaining performance within 5% of the optimal model.

Conclusions:

• Color features and clinical data (hour of life) are most important for jaundice assessment.
• The abdomen provided the most informative skin signals.
• The proposed algorithm offers a promising, automated solution for timely jaundice assessment.