Linda K. Goodwin

Increasing sensitivity of preterm birth by changing rule strengths

We studied two prenatal data sets and two other medical data sets. Our objective was to increase sensitivity (accuracy of preterm birth) by changing the rule strength for the preterm birth class. Two criteria for choosing the optimal rule strength are discussed: the greatest difference between the true-positive and false-positive probabilities and the maximum profit.

Handling missing attribute values in preterm birth data sets

The objective of our research was to find the best approach to handle missing attribute values in data sets describing preterm birth provided by the Duke University. Five strategies were used for filling in missing attribute values, based on most common values and closest fit for symbolic attributes, averages for numerical attributes, and a special approach to induce only certain rules from specified information using the MLEM2 approach. The final conclusion is that the best strategy was to use the global most common method for symbolic attributes and the global average method for numerical attributes.

Data mining methods find demographic predictors of preterm birth.

BackgroundPreterm births in the United States increased from 11.0% to 11.4% between 1996 and 1997; they continue to be a complex healthcare problem in the United States. ObjectiveThe objective of this research was to compare traditional statistical methods with emerging new methods called data mining or knowledge discovery in databases in identifying accurate predictors of preterm births. MethodAn ethnically diverse sample (N = 19,970) of pregnant women provided data (1,622 variables) for new methods of analysis. Preterm birth predictors were evaluated using traditional statistical and newer data mining analyses. ResultsSeven demographic variables (maternal age and binary coding for county of residence, education, marital status, payer source, race, and religion) yielded a .72 area under the curve using Receiving Operating Characteristic curves to test predictive accuracy. The addition of hundreds of other variables added only a .03 to the area under the curve. ConclusionSimilar results across data mining methods suggest that results are data-driven and not method-dependent, and that demographic variables offer a small set of parsimonious variables with reasonable accuracy in predicting preterm birth outcomes in a racially diverse population.

Data mining for preterm birth prediction

Accurate assessment of preterm birth risk remains difficult due to a complex and disorganized knowledge domain, data and information overload, and the absence of reliable and valid tools to measure and predict preterm birth risk. The most persistent limitation for preterm birth risk prediction is our continued lack of understanding about the causes of preterm birth. The purpose of this study was to develop tools and techniques to help better understand the causes of premature birth. Results found only small differences in performance between five different modeling techniques that used neural networks, logistic regression, CART, and software, called PVRuleMiner and FactMiner, specially developed for dealing with problems inherent in clinical data. Contrary to clinical wisdom and earlier studies, most of the predictive power in the database used for this study (1,233 variables total) was found in 32 demographic variables, with only very slight improvements in predictive accuracy when hundreds of variables were added to the models. The ultimate goal of this research is to provide decision support for perinatal care providers to accurately identify patients at risk and assist them with modifiable preterm birth risk factors.

A Closest Fit Approach to Missing Attribute VAlues in Preterm Birth Data

In real-life data, in general, many attribute values are missing. Therefore, rule induction requires preprocessing, where missing attribute values are replaced by appropriate values. The rule induction method used in our research is based on rough set theory.

Coping With Missing Attribute Values Based on Closest Fit in Preterm Birth Data: A Rough Set Approach

Data mining is frequently applied to data sets with missing attribute values. A new approach to missing attribute values, called closest fit, is introduced in this paper. In this approach, for a given case (example) with a missing attribute value we search for another case that is as similar as possible to the given case. Cases can be considered as vectors of attribute values. The search is for the case that has as many as possible identical attribute values for symbolic attributes, or as the smallest possible value differences for numerical attributes. There are two possible ways to conduct a search: within the same class (concept) as the case with the missing attribute values, or for the entire set of all cases. For comparison, we also experimented with another approach to missing attribute values, where the missing values are replaced by the most common value of the attribute for symbolic attributes or by the average value for numerical attributes. All algorithms were implemented in the system OOMIS. Our experiments were performed on the preterm birth data sets provided by the Duke University Medical Center.

A comparison of three closest fit approaches to missing attribute values in preterm birth data

One of the main problems of data mining is imperfection of input data. Such data may be uncertain, vague, and incomplete. In our data set, describing preterm birth, many attribute values were missing, that is, the input data set was incomplete. The main approach to solving the missing attribute value problem was based on a closest fit: a missing attribute value in a case was replaced by the existing attribute value in the best candidate, a case that fits as closely as possible (resembles the most) the case with the missing attribute value. We experimented with three methods based on the idea of the closest fit: looking for the best candidate among the set of all cases, among the cases that belong to the same concept (cases within the same class as the case with missing attribute values), and a special method, where the set of all attributes was restricted to a single attribute with the missing attribute value. In the last method, the missing attribute value was replaced by the most common value within the concept for symbolic attributes, and by the average value of all attribute values of the same concept for numerical attributes.

Preterm Birth Risk Assessed by a New Method of Classifikation Using Selective Partial Matching

In the United States, 8–12% of all newborns are delivered preterm, i.e., before 37 weeks of gestation. Most existing methods to assess preterm birth are based on risk scoring. These methods are only between 17% and 38% predictive in determining preterm birth. Hence there is need for data mining and knowledge discovery in database for predicting birth outcomes in pregnant women. This paper presents a new approach to classification (diagnosis) using selective partial matching. It is shown that our approach is more stable and, in general, more accurate than the method used so far. Our other result shows that classification based on more specific rules is worse.

A comparison of rough set strategies for pre-term birth data

In many applications data are inconsistent: for two distinct cases attribute values are the same but the decisions are different. For example, two patients are characterized identically by all tests, demographic variables, etc., however, one delivers a baby prematurely and the other delivers at full term.Our main objective was to increase sensitivity, i.e., a conditional probability of true positives. In order to achieve this objective we changed a strength multiplier for rules describing preterm cases. In our experiments we induced rules from preterm birth data using LERS (Learning from Examples using Rough Sets). The problem was to make the best use of certain and possible rule sets induced by LERS from inconsistent data. To solve this problem we used eight different strategies: using only certain rules, only possible rules, first certain then possible rules, and both rule sets; combined with two different ways to use rules in complete and partial matching.