A burn injury is a disastrous trauma and can have very wide ranging impacts, including individual, family, and social. Burns patients generally have a long period of hospital stay whose accurate prediction can not only facilitate allocations of scarce medical resources but also help clinicians to counsel patients and relatives at an early stage of care. Besides prediction accuracy, prediction timing of length of hospital stay (LOS) for burn patients is also critical. Early prediction has profound effects on more efficient and effective medical resource allocations and better patient care and management. Hence, the objective of this study is to apply a backpropagation neural network (BPNN) for predicting length of hospital stay (LOS) for burn patients at early stages of care. Specifically, we defined two early-prediction timing, including admission and initial treatment stages. Prediction timing at the admission stage is to predict a burn patient¡¦s LOS when the patient is admitted into the Burns Unit. Prediction at the initial treatment stage refers to the timing right after the first surgery for burn wound excision and skin graft is performed (typically within 72 hours of injury if the patient¡¦s condition allows). Experimentally, we evaluated the prediction accuracy of these two stages, using that achieved at the post-treatment stage (referring to the timing when all surgeries for burn wound excision and skin graft are performed) as benchmarks. The evaluation results showed that prediction LOS at the admission and the initial treatment stages could attain an accuracy of 50.37% and 57.22%, respectively. Compared to the accuracy of 62.13% achieved by the post-treatment stage, the performance reached by the initial treatment stage would consider satisfactory.

Recent advancements in machine learning and the widespread adoption of electronic healthrecords have enabled breakthroughs for several predictive modelling tasks in health care. One such task that has seen considerable improvements brought by deep neural networks is length of stay (LOS) prediction, in which research has mainly focused on adult patients in the intensive care unit. This thesis uses multivariate time series extracted from the publicly available Medical Information Mart for Intensive Care III database to explore the potential of deep learning for classifying the remaining LOS of newborns in the neonatal intensive care unit (NICU) at each hour of the stay. To investigate this, this thesis describes experiments conducted with various deep learning models, including long short-term memory cells, gated recurrentunits, fully-convolutional networks and several composite networks. This work demonstrates that modelling the remaining LOS of newborns in the NICU as a multivariate time series classification problem naturally facilitates repeated predictions over time as the stay progresses and enables advanced deep learning models to outperform a multinomial logistic regression baseline trained on hand-crafted features. Moreover, it shows the importance of the newborn’s gestational age and binary masks indicating missing values as variables for predicting the remaining LOS. / Framstegen inom maskininlärning och det utbredda införandet av elektroniska hälsoregister har möjliggjort genombrott för flera prediktiva modelleringsuppgifter inom sjukvården. En sådan uppgift som har sett betydande förbättringar förknippade med djupa neurala nätverk är förutsägelsens av vistelsetid på sjukhus, men forskningen har främst inriktats på vuxna patienter i intensivvården. Den här avhandlingen använder multivariata tidsserier extraherade från den offentligt tillgängliga databasen Medical Information Mart for Intensive Care III för att undersöka potentialen för djup inlärning att klassificera återstående vistelsetid för nyfödda i den neonatala intensivvårdsavdelningen (neonatal-IVA) vid varje timme av vistelsen. Denna avhandling beskriver experiment genomförda med olika djupinlärningsmodeller, inklusive longshort-term memory, gated recurrent units, fully-convolutional networks och flera sammansatta nätverk. Detta arbete visar att modellering av återstående vistelsetid för nyfödda i neonatal-IVA som ett multivariat tidsserieklassificeringsproblem på ett naturligt sätt underlättar upprepade förutsägelser över tid och gör det möjligt för avancerade djupa inlärningsmodeller att överträffaen multinomial logistisk regressionsbaslinje tränad på handgjorda funktioner. Dessutom visar det vikten av den nyfödda graviditetsåldern och binära masker som indikerar saknade värden som variabler för att förutsäga den återstående vistelsetiden.