| الملخص: | <p>Hypotension is both a dangerous condition and a critical marker for the progression of certain conditions and diseases.Additional research has shown that the more quickly a hypotensive state is predicted, the better the probability of survival.</p><p>This thesis explores the creation of models to predict an acute hypotensive episode at a minimum of one hour before occurrence. Models were based on data acquired from MIMIC II, a large-scale ICU database. A variety of techniques including Parzen models of normality, logistic regression, and neural networks are used to analyse the relationship between vital signs and the occurrence of hypotension in the general ICU population.</p><p>Both univariate and multivariate models were evaluated, and multi-variate neural network models were found to be the best predictors of hypotension in our dataset. The best model was a multi-layer perceptron constructed using values and first-order statistical functions of a patient's mean arterial blood pressure, respiration rate, blood oxygenation level, and heart rate. Best performance occurred at a time of 60 minutes prior to the onset of a patient's first recorded hypotensive episode, and yielded an AUC of 0.84. This corresponded to a classification accuracy of 82% in a balanced dataset and 84% in an unbalanced dataset which more closely rejects the prevalence of hypotensive data in an ICU. These results indicate that there are early patterns underlying general hypotension detectable in patient physiology.</p>
|
|---|