Evolution-assisted discovery of sentinel features in epidemiologic surveillance

John H. Holmes

doi:10.17918/00000981

The use of a genetics-based classifier system (CS) in generating epidemiologic hypotheses was investigated. In addition, epidemiologic analytical techniques were used to evaluate the performance of a CS in this problem domain. Five component studies were implemented, using epidemiologic surveillance data over a range of prevalences. The evaluation study investigated the use of the area under the receiver operating characteristic curve ([theta]) as an alternative to crude accuracy (CA) during the training period. The classification study examined the ability of the CS to classify unencountered patients. The reproducibility study demonstrated the stochastic processes underlying CS performance during training and testing. The payoff-penalty parameterization study investigated the effects of differential penalty for false negative and false positive decisions on learning rate and classification ability. The risk assessment study examined the ability of a CS to derive estimates of risk for purposes of classification. At 50% prevalence, [theta] was identical to CA over the entire training period; with decreasing prevalence, CA increasingly overestimated the learning rate, while [theta] provided more accurate depictions of this measure. Across all four prevalences investigated, the CS was able to classify unseen patients well, with [theta]s ranging from 0.95 at 50% prevalence to 0.78 at 10%. The classifier populations after training indicated considerable generalization; decision rules were discernible on visual examination. When trained and tested using 1,000 different data sets drawn from the same pool, the CS was fairly consistent in terms of learning rate and classification ability, although with sufficient variation to warrant investigating the use of bootstrapping techniques. Biasing the ratio of false positive to false negative (FP:FN) decisions affected the learning rate relative to prevalence. Learning rate was most enhanced at 25% and 10% prevalence by a FP:FN ratio of 4:1 and 10:1, respectively. Across all four prevalences, the CS was able to produce risk estimates that consistently outperformed decision rules derived using logistic regression. The CS was shown to be a useful adjunct to hypothesis generation during epidemiologic surveillance.

Evolution-assisted discovery of sentinel features in epidemiologic surveillance

Files and links (1)

Abstract

Metrics

Details

Evolution-assisted discovery of sentinel features in epidemiologic surveillance

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media