Imputation of missing values by integrating artificial neural networks and case-based reasoning

Does Bayes have it? Decision Support Systems in diagnostic medicine

“…That is, important connections for determining a correct diagnosis are reinforced, and irrelevant connections are attenuated with each training case fed into the system. ANNs have shown great promise in the successful prognosis (Ennett, 2003;Frize et al, 2006Frize et al, , 1995Catley et al, 2006) and in the diagnosis of medical conditions from patient symptom characteristics, for example, in the differentiation of malignant from benign tumors in medical imaging (Goggin et al, 2007). ANNs represent a complex mathematical computation between inputs and the weights in the model, but in the last decade, several methods to extract the weights of the input variables at peak performance were developed, allowing researchers to determine the minimum set of variables leading to the outcome of interest (Rybchynski, 2005;Frize et al, 2006).…”

Predicting Probability of Mortality in the Neonatal Intensive Care Unit

“…The original database contained missing values in patient physiologic variables. Since an ANN cannot process missing values, the incomplete values were imputed [6]. These artificially imputed values could approximate the true clinical data and allow development of ANN mortality prediction models.…”

“…Thirteen variables, a set of mortality risk indicators developed by Ennett in MIRG's previous research [6], were employed as neural network inputs. These variables include lowest pO2/FiO2 ratio, lowest urine output, lowest serum pH, Apgar score at five minutes, lowest platelet count, small for gestational age status, highest sodium, highest respiratory rate, highest pCO2, birth weight, lowest glucose, lowest temperature, and highest mean blood pressure.…”

“…classifying survival and death cases directly). Ideally, to ensure confidence of using a model for such a scenario, a specificity value of close to 100% and a sensitivity value of at least 50% would be acceptable by users [6]. For the development of a mortality-screening tool without affecting its probability estimation, we moved the value of the cutoff point between 0 and 0.5 to compare the resulting sensitivity and specificity.…”

“…Previous studies by the members of the Medical Information technology Research Group (MIRG) showed that feed-forward neural networks were successfully used to predict NICU mortality [5]. The performance of neural network mortality estimation was comparable to or outperformed the results of the 'SNAP' (Score for Neonatal Acute Physiology) scoring system or fuzzy-logic classifiers [6,7], etc. A new set of 13 variables was developed through ANN learning as the important predictors of neonatal mortality [6].…”