Deep learning-based neonatal outcome prediction: an LSTM autoencoder framework for pattern analysis and risk assessment.

The phenomenon of neonatal mortality is a significant issue in health care, which necessitates the use of advanced analytical tools for early risk prediction. This work utilized the Medical Information Mart (MIMIC) Pediatric Intensive Care (PIC) database to develop a framework for identifying key biomarkers associated with mortality risk. The combination of many clinical data sources comprising more than two thousand Neonatal Intensive Care Unit (NICU) Neonatal patient admissions (both alive and dead) were extracted. Biomarkers with an abnormal value were taken into consideration for further analysis. The patterns were analyzed using Long Short-Term Memory (LSTM) autoencoders that converted the biomarker data of both deceased and live neonates into meaningful representations. Similar biomarker patterns were clustered to gain insights into the mortality risk factors. The suggested framework is found to be superior to the existing neonatal scoring systems with a sensitivity of 88.1%, specificity of 89.5%, and Area Under Receiver Operating Characteristic Curve (AUCROC) of 0.92. These results indicate the possibility of employing the biomarker patterns as predictors of neonatal mortality and the basis of a decision-support model.