Machine learning approach to analyzing complex care coordination patterns for medically complex children.

backgroundThe number of children with medical complexity (CMC) in Japan has exceeded 20,000, doubling over the past 15 years, forming a significant population group internationally. While care coordination for CMC requires consideration of growth and school attendance, the relatively small population size compared to elderly care makes it challenging for care coordinators to gain sufficient experience. We investigated whether machine learning models could assist in care coordination for CMC.

methodsWe developed three tree-based gradient boosting prediction models using 1,042 care sheets from a pediatric respite facility. The first model estimated ventilator use to verify if care patterns could predict symptoms; the second predicted home doctor utilization to examine if symptoms could indicate required care; and the third estimated age to analyze school-related care needs. SHAP values were used to interpret feature importance in the prediction models. Each model was independently created and validated ten times to account for potential sampling bias among respite facility users.

resultsThe highest F1 scores for the three models were 0.83, 0.826, and 0.829, respectively. The ventilator model identified equipment maintenance services as highly significant; the home doctor model revealed patterns in complex service utilization; and the age model demonstrated varying important factors across school-age transitions.

conclusionThis study highlights AI's potential in identifying crucial factors related to CMC services and its effectiveness as a valuable tool for service coordination. Despite the limited dataset of 1,042 cases, our findings emphasize the importance of digital preservation of CMC healthcare information.

trial registrationRetrospectively registered. CLINICAL TRIAL NUMBER: Not applicable.