Neural network detection of files of suicidal patients and suicidal profiles

The optimal configuration of backpropagation (BP) neural networks was determined after 35 trials with different BP configurations evaluating the total detection rate. Ten different training and testing sets were used to identify optimal samples. All trials included sample files of patients with medically serious suicidal attempts (MSSA) and those of non-suicidal patients. Fifty files were used in each group for training and 49 files for testing with no overlap between the samples. The target variable for training was seriousness of suicide attempt (0 = non-suicidal, 1 = MSSA). The input set included 44 demographic, clinical and patient-history variables. The optimal results showed that 93.8% of MSSA and 89.8% of the non-suicidal patient files were detected. Total success rate (TSR) was 91.8% and positive and negative prediction values (PPV, NPV) were 92% and 95.6%, respectively. Living alone (6.76), treatment compliance (5.86), drug abuse or dependence (2.8), global assessment of functioning (GAF) score (1.49), non-paranoid delusions (1.22) and suicide of first degree relative (1.1) were highly associated with MSSA according to the Garson calculation. However, logistic regression attributed high importance to hallucinations (p < 0.0001), diagnosis (p < 0.002), number of children (p < 0.006), GAF score (p < 0.006), employment status (p < 0.02) and stressors (p < 0.03). It was shown that: backpropagation neural networks are very successful in identifying records of MSSA patients; a high GAF score is associated with high risk of MSSA and is the only common variable identified by both methods; and backpropagation identified two non-specific factors (living alone and treatment compliance) whereas statistics found specific factors (hallucinations and diagnosis) highly associated with MSSA.