Application for monitoring the patients with suicidal tendencies
A. C. Bratan, V. Tau, E. D. Franti, C. Niculae, A. V. Tebeanu, M. Dascalu, A. CiobanuIntroduction
The application present into this paper was developed during a collaboration with two psychiatric hospitals from Romania, and it is new because it is based on data obtained from patients who attempted suicide and survived against their will, and thus they provide to the doctors many details about what they felt during the moments when they tried to kill themselves. The recordings from the patients’ interviews showed that at the moments when they tried to kill themselves, the patients experienced intense emotions specific to suicide, that remained impregnated for many hours or even days in their voices.
Objectives
The application was developed with a Convolutional Neural Network (CNN) that was trained to identify the specific pattern for each level of intensity of the emotions specific to suicide, imprinted in the voices of three categories of patients:
first category included 49 patients that just attempted suicide and survive (they were selected only patients with NO mental issues); these patients had strong emotions specific to suicide; second category included 31 patients who were treated in the hospital during 4 weeks (after they survived to suicide attempts); these patients had emotions specific to suicide with medium intensity; third category included 49 patients with depression and suicidal ideation (meaning that they had only thought about suicide, without implementing any concrete action to kill themself
); these patients had emotions specific to suicide with low intensity.
Methods
The patients were informed about the study objectives, stages, and data protection measures, and were assured of their right to withdraw at any time. All subjects gave informed consent and assent for inclusion before participating in the study. The patients’ interviews were audio recorded using two Sony ICD-PX240 recorders and were saved in MP3 format.
Results
The CNN achieved an 81.6% accuracy rate in classification, the evolution of the accuracy parameter during the 10 epochs.
Image 1: long description.