Predicting death by suicide using administrative health care system data: Can recurrent neural network, one-dimensional convolutional neural network, and gradient boosted trees models improve prediction performance?

BACKGROUND

Suicide is a leading cause of death, particularly in younger persons, and this results in tremendous years of life lost.

OBJECTIVE

To compare the performance of recurrent neural networks, one-dimensional convolutional neural networks, and gradient boosted trees, with logistic regression and feedforward neural networks.

METHODS

The modeling dataset contained 3548 persons that died by suicide and 35,480 persons that did not die by suicide between 2000 and 2016. 101 predictors were selected, and these were assembled for each of the 40 quarters (10 years) prior to the quarter of death, resulting in 4040 predictors in total for each person. Model configurations were evaluated using 10-fold cross-validation.

RESULTS

The optimal recurrent neural network model configuration (AUC: 0.8407), one-dimensional convolutional neural network configuration (AUC: 0.8419), and XGB model configuration (AUC: 0.8493) all outperformed logistic regression (AUC: 0.8179). In addition to superior discrimination, the optimal XGB model configuration also achieved superior calibration.

CONCLUSIONS

Although the models developed in this study showed promise, further research is needed to determine the performance limits of statistical and machine learning models that quantify suicide risk, and to develop prediction models optimized for implementation in clinical settings. It appears that the XGB model class is the most promising in terms of discrimination, calibration, and computational expense.

LIMITATIONS

Many important predictors are not available in administrative data and this likely places a limit on how well prediction models developed with administrative data can perform.