Predicting death by suicide following an emergency department visit for parasuicide with administrative health care system data and machine learning.

BACKGROUND: Suicide is a leading cause of death worldwide and results in a large number of person years of life lost. There is an opportunity to evaluate whether administrative health care system data and machine learning can quantify suicide risk in a clinical setting. METHODS: The objective was to compare the performance of prediction models that quantify the risk of death by suicide within 90 days of an ED visit for parasuicide with predictors available in administrative health care system data.The modeling dataset was assembled from 5 administrative health care data systems. The data systems contained nearly all of the physician visits, ambulatory care visits, inpatient hospitalizations, and community pharmacy dispenses, of nearly the entire 4.07 million persons in Alberta, Canada. 101 predictors were selected, and these were assembled for each of the 8 quarters (2 years) prior to the quarter of death, resulting in 808 predictors in total for each person. Prediction model performance was validated with 10-fold cross-validation. FINDINGS: The optimal gradient boosted trees prediction model achieved promising discrimination (AUC: 0.88) and calibration that could lead to clinical applications. The 5 most important predictors in the optimal gradient boosted trees model each came from a different administrative health care data system. INTERPRETATION: The combination of predictors from multiple administrative data systems and the combination of personal and ecologic predictors resulted in promising prediction performance. Further research is needed to develop prediction models optimized for implementation in clinical settings. FUNDING: There was no funding for this study.

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.

Major depressive episodes and mortality in the Canadian household population.

OBJECTIVE: To explore the association between major depressive episodes (MDE) and subsequent mortality in a representative sample of the general household population, with adjustment for other determinants of mortality. METHOD: The analysis used four datasets from the Canadian Community Health Survey (CCHS); the CCHS 1.1 (conducted in 2000 and 2001), the CCHS 1.2 (conducted in 2002), the CCHS 2.1 (conducted in 2003 and 2004) and the CCHS 3.1 (conducted in 2005 and 2006). Each survey included an assessment of past-year major depressive episodes (MDEs) and was linked to mortality data from the Canadian Mortality Database for January 1, 2000 to December 31, 2011. The hazard ratio (HR) for all-cause mortality was estimated in each survey sample. Random effects, individual-level meta-analysis was used to pool estimates from the four survey data sets. Estimates were adjusted for other determinants of mortality prior to pooling in order to help quantify the independent contribution of MDE to all-cause mortality. RESULTS: The unadjusted HR was 0.77 (95% CI 0.63-0.95). A naïve interpretation of this HR suggests a protective effect of MDE, but the estimate was found to be strongly confounded by age (age adjusted HR for MDE: 1.61, 95% CI 1. 34-1.93) and by sex (sex adjusted HR for MDE: 1.15, 95% CI 0.75-1.77). The age and sex adjusted HR was: 1.70 (95% CI 1.45-2.00). No evidence of effect modification by any determinant of mortality was found, including sex. After adjustment for a set of mortality risk factors, the pooled HR was weakened, but remained statistically significant, HR = 1.29 (I-squared = < 1%, tau-squared < 0.001, 95% CI 1.10-1.51). Smoking was the strongest single confounding variable. CONCLUSIONS: MDE is associated with elevated mortality. The elevated risk is partially attributable to psychosocial, behavioral and health-related determinants. Since MDE itself may have caused changes to these variables, these estimates cannot fully quantify the independent contribution of MDE to mortality. However, these results suggest that clinical and public health efforts to counteract the effect of MDE on mortality may benefit from attention to a broad set of mortality risk factors e.g. smoking, physical activity, management of medical conditions.

Validating screening tools for depression in stroke and transient ischemic attack patients.

OBJECTIVE: The best screening questionnaires for detecting post-stroke depression have not been identified. We aimed to validate four commonly used depression screening tools in stroke and transient ischemic attack patients. METHODS: Consecutive stroke and transient ischemic attack patients visiting an outpatient stroke clinic in Calgary, Alberta (Canada) completed a demographic questionnaire and four depression screening tools: Patient Health Questionnaire (PHQ)-9, PHQ-2, Hospital Anxiety and Depression Scale (HADS-D), and Geriatric Depression Scale (GDS-15). Participants then completed the Structured Clinical Interview for DSM-IV (SCID), the gold-standard for diagnosing major depression. The questionnaires were validated against the SCID and sensitivity and specificity were calculated at various cut-points. Optimal cut-points for each questionnaire were determined using receiver-operating curve analyses. RESULTS: Among 122 participants, 59.5% were diagnosed with stroke and 40.5% with transient ischemic attack. The point prevalence of SCID-diagnosed current major depression was 9.8%. At the optimal cut-points, the sensitivity and specificity for each screening tool were as follows: PHQ-9 (sensitivity: 81.8%, specificity: 97.1%), PHQ-2 (sensitivity: 75.0%, specificity: 96.3%), HADS-D (sensitivity: 63.6%, specificity: 98.1%), and GDS-15 (sensitivity: 45.5%, specificity: 84.8%). Areas under the receiver operating characteristic curves were as follows: PHQ-9 86.6%, PHQ-2 86.7%, HADS-D 85.9%, and GDS-15 66.3%. CONCLUSIONS: The PHQ-2 and PHQ-9 are both suitable depression screening tools, taking less than 5 minutes to complete. The HADS-D does not appear to have any advantage over the PHQ-based scales, even though it was designed specifically for medically ill populations. The GDS-15 cannot be recommended for general use in a stroke clinic based on this study as it had worse discrimination due to low sensitivity.