Machine Learning Model Predicts One-Year Rehospitalization in Preterm Infants

Predicting the Fragile Transition from NICU to Home

Preterm birth remains a significant driver of pediatric morbidity, requiring complex transitions from the neonatal intensive care unit (NICU) to home. While late preterm infants constitute the majority of these births, they remain a vulnerable population with high rates of unplanned readmission [1]. Beyond immediate clinical complications, these infants face long-term risks including respiratory impairment and developmental delays, particularly when conditions like bronchopulmonary dysplasia are present [2]. Socioeconomic disparities further complicate postnatal growth and recovery, often exacerbating health inequities in the first year of life [3]. Although primary care plays a vital role in mitigating these risks, clinicians often lack precise tools to identify which infants require the most intensive post-discharge monitoring [4]. To address this gap, researchers recently evaluated whether machine learning models applied to neonatal clinical data can accurately flag infants at the highest risk for rehospitalization, potentially allowing pediatricians to target early interventions and closer outpatient surveillance.

Quantifying the Burden of Post-Discharge Readmission

The transition from the neonatal intensive care unit to the outpatient setting represents a period of profound physiological vulnerability for preterm infants. To quantify the scale of this risk, researchers conducted a retrospective cohort study tracking secondary health encounters among 2,226 preterm infants. The findings revealed that rehospitalization within one year occurred in 358 of the 2,226 infants (16.1%), highlighting a substantial subset of patients who require acute care intervention shortly after their initial discharge. For practicing pediatricians and family physicians, this 16.1% readmission rate serves as a critical benchmark for evaluating the adequacy of outpatient support systems and discharge planning. The temporal distribution of these readmissions points to a particularly narrow window of high risk. The study determined that one-third of all rehospitalizations occurred within 30 days of discharge. Because the first month at home is the most precarious phase for this population, identifying the specific clinical and socioeconomic drivers behind these early readmissions is essential for coordinating targeted post-discharge surveillance and parental education.

Cohort Characteristics and Data Integration

To develop a predictive framework for these post-discharge outcomes, the investigators analyzed a retrospective cohort of 2,226 preterm infants born between 2018 and 2023 at a tertiary-care pediatric hospital in Israel. This setting allowed for the consistent collection of high-fidelity clinical data across a diverse range of gestational ages and medical complexities. The researchers constructed their dataset by integrating information from multiple institutional sources, primarily drawing from neonatal intensive care unit and inpatient records. This comprehensive data integration captured the full clinical history of each infant, aggregating laboratory results and hospitalization outcomes alongside longitudinal clinical data. By mapping the physiological and therapeutic variables that preceded discharge, the authors built a foundation for the eXtreme Gradient Boosting algorithm (XGBoost). XGBoost is a machine learning method that builds a predictive model by sequentially combining the strengths of multiple simpler decision trees to improve overall accuracy. In a clinical context, this approach allows the algorithm to weigh dozens of interacting patient variables simultaneously, moving beyond single-risk-factor assessments to generate a personalized risk profile for each infant.

Algorithmic Performance and Predictive Accuracy

The primary objective of the study was to train this machine learning model to predict rehospitalization within one year of discharge using 20 clinical predictors. When evaluated using standard diagnostic metrics, the predictive model achieved an Area Under the Curve (AUC) of 0.69 (95% CI: 0.59 to 0.78). The AUC is a statistical measure representing the model's overall ability to correctly classify infants into rehospitalized and non-rehospitalized groups, with 0.69 indicating moderate discriminative power. The model demonstrated a high specificity of 0.87, meaning it was highly effective at identifying infants who were not at risk for readmission. This was mirrored by a negative predictive value of 0.87, suggesting that when the algorithm flags an infant as low-risk, that assessment is accurate in 87% of cases. However, the model showed more limited performance in identifying all true positive cases, yielding a sensitivity of 0.38. This lower sensitivity indicates that the algorithm missed a significant portion of infants who were eventually rehospitalized. Furthermore, the positive predictive value was 0.38, meaning that only 38% of the infants flagged as high-risk actually required readmission within the first year. For the practicing clinician, these metrics suggest that while the model is a reliable tool for ruling out readmission and identifying low-risk infants who may safely require less intensive surveillance, its current iteration cannot replace clinical vigilance for capturing every infant at risk for morbidity.

Identifying the High-Risk Clinical Profile

Beyond overall predictive accuracy, the algorithm identified specific clinical characteristics strongly associated with increased rehospitalization risk. Among the primary physiological drivers, the researchers identified early gestational age and lower birth weight as key predictors. Interestingly, the model also flagged a discharge weight greater than 2000 g as a significant predictor of one-year rehospitalization. This finding suggests that reaching a standard weight milestone for discharge does not necessarily mitigate the underlying vulnerabilities associated with severe prematurity. The duration of initial medical intervention also played a major role, as a prolonged neonatal intensive care unit stay was strongly linked to a higher likelihood of returning to the hospital. The presence of specific comorbidities further refined the high-risk profile. The model identified trisomy and the presence of gastrointestinal and neurological conditions as critical predictors for readmission. Respiratory health was another major factor, with bronchopulmonary dysplasia (a chronic lung disease common in premature infants requiring prolonged mechanical ventilation or oxygen therapy) significantly increasing risk. The complexity of the initial hospitalization also mattered, as surgical interventions and abnormal laboratory values during the neonatal period were identified as key markers of future instability. Finally, the study incorporated socioeconomic data, identifying a low socioeconomic score as a key predictor. This highlights how environmental and financial stressors directly impact post-discharge outcomes. By integrating these diverse clinical and social markers, the model provides a multidimensional view of infant vulnerability, offering a framework that could eventually help pediatricians allocate targeted early interventions and intensive follow-up care to the infants who need it most.

References

1. Joyner JA, Papermaster A, Champion J. Characteristics of late preterm infant readmissions: A systematic review. Journal of the American Association of Nurse Practitioners. 2024. doi:10.1097/JXX.0000000000000986

2. Almutairi M, Alomari S, Alomari N, et al. Upper Airway Manifestations and Otolaryngologic Management of Bronchopulmonary Dysplasia in Preterm Infants: A Systematic Review. ENT Updates. 2025. doi:10.54963/entu.v15i3.1256

3. Ravi K, Young A, Beattie RM, Johnson MJ. Socioeconomic disparities in the postnatal growth of preterm infants: a systematic review. Pediatric Research. 2024. doi:10.1038/s41390-024-03384-0

4. Starfield B, Shi L, Macinko J. Contribution of Primary Care to Health Systems and Health. Milbank Quarterly. 2005. doi:10.1111/j.1468-0009.2005.00409.x