LONG SHORT-TERM MEMORY NETWORK FOR PREDICTING HYPOTENSIVE EPISODES IN AKI PATIENTS UNDERGOING CKRT WITH TRANSFER LEARNING APPROACH FROM ADULT TO PEDIATRIC PATIENTS

Critically ill pediatric patients undergoing Continuous Kidney Replacement Therapy (CKRT) are at high risk of adverse hemodynamic events, including Acute Hypotensive Episodes (AHEs), which are the leading causes of morbidity and mortality in the Pediatric Intensive Care Unit (PICU). Despite their clinical impact, no validated predictive models currently exist for early detection of AHE in children, mainly due to the rarity of CKRT in pediatrics and the limited availability of high-quality time-series data. Identifying early markers of AHE could provide crucial insights into patient hemodynamic stability and guide timely interventions to prevent life-threatening complications.

We developed a deep learning model based on a Long Short-Term Memory (LSTM) network to predict AHEs from multivariate physiological time-series (arterial pressure, heart rate, oxygen saturation, and respiratory rate). To overcome data scarcity, we implemented a transfer learning strategy: the model was pre-trained on the large MIMIC-III adult ICU dataset (split into 70% patients for training set and 15% patients for validation and test set). The training set, consisting of 1934 patients, contained 4249 positive data samples and 144398 negative data samples. Subsequently the model was fine-tuned on a smaller, monocentric pediatric cohort (14 CKRT patients) from the Giannina Gaslini Institute (Genoa, Italy). The pediatric dataset (training set of 9 patients, contained 26 positive and 740 negative data samples.) included retrospective and prospective data, harmonized and filtered through a standardized preprocessing pipeline ensuring temporal consistency and quality control. Performance was assessed using multiple metrics including AUROC, precision, recall, and specificity. This research was funded by the Italian Ministry of Health (Ricerca Finalizzata 2019)

The baseline model trained on MIMIC-III achieved an AUROC of 0.926, which increased to 0.934 after transfer learning on the Gaslini cohort, indicating excellent generalization from adult to pediatric populations. Despite the strong class imbalance (≈3% positive sequences), the model maintained a high recall (0.90) and acceptable specificity (0.88), identifying the majority of hypotensive events while minimizing missed episodes. Precision remained moderate (0.20) due to the scarcity of positive events, but this trade-off is clinically acceptable for an early-warning tool in critical care. The model successfully anticipated AHE onset up to five minutes in advance, offering a feasible window for clinical intervention.

This study demonstrates, for the first time, the successful application of transfer learning from adult to pediatric ICU data for predicting acute hypotensive episodes in children receiving CKRT. The approach effectively bridges the gap between data-rich adult datasets and data-limited pediatric contexts, enabling robust prediction in a rare and high-risk clinical population. Beyond its technical innovation, the model has strong clinical implications: by providing an early signal of hemodynamic instability, it could support individualized fluid and vasoactive management, reduce adverse events, and ultimately improve outcomes in pediatric critical care. Future multicentric studies with larger pediatric cohorts are warranted to validate and further refine this predictive framework.