Assessment of individual risk of death using self-report data: an artificial neural network compared with a frailty index

Abstract

Objectives: To evaluate the potential of an artificial neural network (ANN) in predicting survival in elderly Canadians, using self-report data.

Design: Cohort study with up to 72 months follow-up.

Setting: Forty self-reported characteristics were obtained from the community sample of the Canadian Study of Health and Aging. An individual frailty index score was calculated as the proportion of deficits experienced. For the ANN, randomly selected participants formed the training sample to derive relationships between the variables and survival and the validation sample to control overfitting. An ANN output was generated for each subject. A separate testing sample was used to evaluate the accuracy of prediction.

Participants: A total of 8,547 Canadians aged 65 to 99, of whom 1,865 died during 72 months of follow-up.

Measurements: The output of an ANN model was compared with an unweighted frailty index in predicting survival patterns using receiver operating characteristic (ROC) curves.

Results: The area under the ROC curve was 86% for the ANN and 62% for the frailty index. At the optimal ROC value, the accuracy of the frailty index was 70.0%. The ANN accuracy rate over 10 simulations in predicting the probability of individual survival mean+/-standard deviation was 79.2+/-0.8%.

Conclusion: An ANN provided more accurate survival classification than an unweighted frailty index. The data suggest that the concept of biological redundancy might be operationalized from health survey data.