A machine learning approach to using ultrasound for body composition and nutritional status assessment in newborns: a pilot study protocol.
Authors
Affiliations (9)
Affiliations (9)
- Department of Engineering, Boston College, Chestnut Hill, MA, 02467, USA. [email protected].
- William F. Connell School of Nursing, Boston College, Chestnut Hill, MA, USA. [email protected].
- William F. Connell School of Nursing, Boston College, Chestnut Hill, MA, USA.
- Department of Engineering, Boston College, Chestnut Hill, MA, 02467, USA.
- Department of Computer Science, Boston College, Chestnut Hill, MA, USA.
- Department of Pediatrics and Child Health, Jimma University, Jimma, Ethiopia.
- Department of Radiology, Boston Children's Hospital, Boston, MA, USA.
- Harvard Medical School, Boston, MA, USA.
- Department of Pediatrics, Brigham & Women's Hospital, Boston, MA, USA.
Abstract
Accurate assessment of infant body composition, specifically fat and fat-free mass, is crucial for evaluating growth and nutritional status. Existing methods, such as air displacement plethysmography and dual-energy X-ray absorptiometry, are expensive, require specialized facilities, and demand trained personnel. Ultrasound offers a promising alternative as a portable, low-cost tool capable of distinguishing adipose tissue from skeletal muscle. However, its feasibility for widespread use in diverse clinical settings remains uncertain. This pilot study aims to assess the feasibility, acceptability, and reliability of an ultrasound scanning protocol for measuring body composition in infants while establishing a foundation for artificial intelligence (AI)-enabled analysis to enable whole-body composition estimation from ultrasound images. We will recruit 50 infants (20 preterm and 30 full-term infants) from two sites: Brigham and Women's Hospital (Boston, MA, USA) and Jimma Medical Center (Jimma, Ethiopia). Feasibility will be assessed through metrics such as recruitment rates, scan completion, and session duration. Acceptability will be measured using clinician feedback, and reliability will be evaluated using intra-class correlation coefficients (ICCs) for ultrasound image acquisition. A comprehensive database of ultrasound images and corresponding body composition metrics will be developed, forming the foundation for training AI models. Preliminary machine learning (ML) models, including convolutional neural networks (CNN), will be developed to predict body composition. The accuracy of these models will be evaluated using metrics such as root mean squared error (RMSE), mean absolute percent error (MAPE), mean absolute error (MAE), and mean squared error (MSE). This study will determine the feasibility of integrating ultrasound-based body composition assessment into neonatal clinical workflows as a potential future application. We will evaluate protocol adherence, scan reliability, and clinician and family acceptability to guide further protocol optimization. Findings will inform the design of a larger-scale study and contribute to refining AI models for clinical use. Ultimately, this approach aims to improve the accessibility, accuracy, and efficiency of body composition assessments, particularly in low-resource settings, where it could enable frontline healthcare workers to perform these assessments without specialized training, improving care for vulnerable infants.