Nursing Process Data for Health Care Cost Prediction Using Machine Learning: Longitudinal Study

Machine learning (ML) has been demonstrated to enhance health care cost prediction by handling high-dimensional data and identifying complex patterns. However, current risk-adjustment models rarely incorporate structured nursing information derived from the nursing process. This information captures care needs and human responses to health problems.

This study aimed to evaluate the impact of integrating nursing process data into ML-based predictive models of individual health care costs, including cost component analyses, compared with models based solely on sociodemographic, clinical, and morbidity-related variables.

A retrospective observational study was conducted using a population-based cohort of 1,691,075 individuals aged 15 years or younger who were registered with the Canary Islands Health Service. Predictors were derived from data available up to 2017 and included sociodemographic and clinical variables, Adjusted Morbidity Groups, health care use, and structured nursing records (Functional Health Patterns [FHP], North American Nursing Diagnosis Association [NANDA], Nursing Outcomes Classification [NOC], and Nursing Interventions Classification [NIC]). Predictive models were developed using feedforward neural networks and extreme gradient boosting; predictions were combined using an ensemble approach. An autoencoder was applied as a dimensionality-reduction technique for the nursing variables. Model performance with and without nursing variables was compared on total cost and individual cost components, and the coefficient of determination (R²) was used on the test set.

Including the nursing methodology yielded small numerical increases in predictive performance. With respect to total cost, the ensemble model improved the R² from 0.5023 to 0.5058 when the nursing variables were added. Although directionally consistent, these gains were limited in magnitude. In the component-level analyses, performance gains were observed in hospital care (R²=0.2396) and pharmaceutical costs (R²=0.6631). Reducing 789 nursing variables to 16 latent dimensions using an autoencoder substantially simplified the feature space, with predictive performance remaining broadly comparable but without a substantial additional gain.

Integrating structured information from the nursing process is associated with small incremental improvements in ML-based predictive models and complements commonly used sociodemographic, clinical, and morbidity variables. The systematic incorporation of nursing data into predictive tools may contribute to more accurate health care cost prediction and support more holistic, person-centered approaches.