Digital twin for case prevention and management in premature neonates

Premature newborns are infants who are born before fully completing their developmental processes in the womb. These infants spend a critical period in the Neonatal Intensive Care Unit (NICU), where their growth and development are closely monitored. During this time, premature infants are highly susceptible to various medical conditions, making it essential to continuously track and record their medical parameters and laboratory findings. Some medical conditions can progress so rapidly that they might lead to the infant's death within hours after the onset. Therefore, early detection before the occurrence of such cases and developing strategies for effective case management are critically important. Digital twins represent physical objects or processes by transferring sensor data into a digital environment, creating virtual representations. Through digital twins, it is also possible to generate artificial scenarios to simulate real-life conditions. Artificial intelligence (AI) enhances digital twin models by enabling forward-looking predictions and inferences. In this thesis, laboratory findings, medical parameters, and hyperspectral images from 47 premature infants treated at Selçuk University's NICU were analyzed. All collected data, including images and parameters, were recorded with timestamps. Based on this data, AI models were developed for forecasting future CRP, bilirubin, and hemoglobin values through time series analyses using current and past medical parameters, as well as disease prediction using timestamped hyperspectral images. Long Short-Term Memory (LSTM), XGBoost, and LSTM-Multi-Head Attention (MHA) mechanisms were utilized to develop time series analysis models for predicting future laboratory findings. The LSTM-based time series analysis model achieved R-squared accuracies of 29.09% for CRP, 36.36% for bilirubin, and 46.63% for hemoglobin predictions. The XGBoost model obtained Mean Squared Errors (MSE) of 9.76 for bilirubin and 16.8 for hemoglobin predictions. Using the LSTM-MHA model, R-squared accuracies of 91.78%, 66.76%, and 77.62% were achieved for CRP, bilirubin, and hemoglobin, respectively. Additionally, a Hyperspectral Vision Transformer (HSViT) model was developed for disease prediction using timestamped hyperspectral imagery. Principal Component Analysis (PCA) was applied to the hypercube dataset at different levels 5, 50, and 100 components and models were trained on these reduced datasets. The best-performing model was trained on a dataset reduced to 50 principal components, achieving an accuracy of 98.92%, precision of 96.48%, sensitivity of 99.50%, and an F1-score of 97.95%. This thesis demonstrates that artificial intelligence models integrated with digital twins can effectively predict future medical parameters and diseases in premature newborns, providing a promising early-warning system for disease prevention and management.