Predicting serum phosphate levels in very preterm infants using machine learning
Åsbjørn S Westvik1,2, Oliver Tomic3, Charlotte Tscherning1,2
1Institute of Clinical Medicine, University of Oslo, Oslo, Norway.
Insights
Small for gestational age (SGA) infants face high hypophosphatemia risk. Machine learning models can estimate serum phosphate levels up to 24 hours in advance, aiding nutritional management in preterm infants.
Area of Science:
- Neonatal Medicine
- Biostatistics
- Machine Learning Applications
Background:
- Hypophosphatemia is prevalent in very preterm and small for gestational age (SGA) infants, leading to severe complications.
- Current diagnosis relies on direct serum or plasma phosphate measurements.
- Machine learning (ML) offers potential for indirect phosphate estimation using routine clinical and blood gas data.
Purpose of the Study:
- To analyze first-week electrolyte concentrations in very preterm infants within the ImNuT-trial.
- To develop and externally validate ML models for estimating serum phosphate levels in the first postnatal week.
- To utilize routinely collected clinical and nutritional data for phosphate estimation in preterm infants.
Main Methods:
- Retrospective analysis of 120 infants (<29 weeks gestation) under a standardized nutritional protocol.
- Description of electrolyte trajectories (calcium, potassium, sodium, phosphate) stratified by hypophosphatemia and SGA status.
- Development of ML models (elastic net, gradient boosting, random forests, etc.) for concurrent and future phosphate prediction (12h, 24h) using 22 candidate predictors and cross-validation, including an external cohort of 40 infants.
Main Results:
- 33.6% of 119 infants developed hypophosphatemia in the first week.
- SGA infants exhibited significantly higher and earlier risk of hypophosphatemia (log-rank p < 0.0001).
- Elastic net models demonstrated best external performance, consistently using calcium, potassium, gestational age, SGA status, and weight z-scores as predictors.
Conclusions:
- SGA status is a critical factor for early hypophosphatemia in very preterm infants.
- ML models can moderately and consistently estimate serum phosphate levels 12-24 hours ahead using available data.
- These ML models may serve as screening tools to optimize nutritional support and reduce blood sampling, pending prospective validation.
Background:
Hypophosphatemia in very preterm and small for gestational age (SGA) infants is common and associated with serious complications. Diagnosis relies on phosphate concentrations, measured in serum or plasma. Machine learning (ML) may enable indirect estimation of serum phosphate based on routinely collected clinical data and results from blood gas analysis.
Objectives:
To (1) describe first week electrolyte trajectories in the ImNuT-trial (Clinicaltrials.gov ID: NCT03555019), and (2) develop and externally validate ML models to estimate serum phosphate levels during the first postnatal week in very preterm infants, using routinely collected clinical and nutritional data.
Methods:
We performed a retrospective analysis of 120 infants born <29 weeks' gestation, all managed under a standardized nutritional protocol. Electrolyte trajectories (calcium, potassium, sodium, phosphate) were described using daily means stratified by hypophosphatemia and SGA status. For ML, we defined three regression tasks: concurrent phosphate prediction at blood gas analysis, and forecasts 12 h and 24 h ahead. Twenty-two candidate predictors from blood gases, anthropometry, SGA status and nutrient intakes were subjected to RENT-based variable selection within leave-one-group-out cross-validation. Multiple algorithms (elastic net, gradient boosting, random forests, k-nearest neighbours, kernel ridge, XGBoost) were trained under three preprocessing strategies (complete cases, complete cases without outliers, imputed data). Performance was evaluated with RMSE and R 2 using nested cross-validation and an independent external cohort of 40 very preterm infants.
Results:
Among 119 infants with phosphate measurements (607 samples), 33.6% developed hypophosphatemia in week one. SGA infants had a markedly higher and earlier risk of hypophosphatemia compared to appropriate-for-gestational-age infants (log-rank p < 0.0001). Six predictors (calcium, potassium, gestational age, SGA status, birth weight z-score, daily weight z-score) were consistently retained across all prediction tasks. Elastic net models performed best externally on all three prediction tasks. Performance was robust across preprocessing strategies.
Conclusions:
SGA status is a major determinant of early hypophosphatemia risk in very preterm infants. ML models using routinely available clinical and nutritional data can moderately but consistently estimate serum phosphate levels, including 12-24 h ahead, and may serve as screening tools to guide nutritional management and reduce unnecessary blood sampling. Prospective validation is warranted before clinical implementation.
