Deep Sequential Models for Suicidal Ideation From Multiple Source Data

Abstract

This article presents a novel method for predicting suicidal ideation from Electronic Health Records (EHR) and Ecological Momentary Assessment (EMA) data using deep sequential models. Both EHR longitudinal data and EMA question forms are defined by asynchronous, variable length, randomly-sampled data sequences. In our method, we model each of them with a Recurrent Neural Network (RNN), and both sequences are aligned by concatenating the hidden state of each of them using temporal marks. Furthermore, we incorporate attention schemes to improve performance in long sequences and time-independent pre-trained schemes to cope with very short sequences. Using a database of 1023 patients, our experimental results show that the addition of EMA records boosts the system recall to predict the suicidal ideation diagnosis from 48.13% obtained exclusively from EHR-based state-of-the-art methods to 67.78%. Additionally, our method provides interpretability through the t-SNE representation of the latent space. Further, the most relevant input features are identified and interpreted medically.

The use of machine learning on administrative and survey data to predict suicidal thoughts and behaviors: a systematic review

“…Among studies that aimed to predict suicidal thoughts using administrative data, Peis et al. ( 17 ) and McKernan et al. ( 18 ) showed that mental illness, related inpatient utilization, and previous suicidal thoughts and attempt(s) are common risk factors in addition to some social factors like shared residence and living with offspring or siblings.…”

“…( 41 ), and Peis et al. ( 17 ) found that depression ( 41 , 48 ), anxiety ( 41 , 48 ), stress ( 41 ), previous suicidal thoughts or suicide attempts ( 17 ), and shared residence or living with siblings or offspring ( 17 ) were risk factors for suicidal thoughts ( Supplementary Table 2 ). Depression ( 45 , 60 , 64 , 65 ), impulsivity ( 60 ), borderline personality disorder ( 69 ), post-traumatic stress disorder ( 69 ), alcohol use disorders identification test (AUDIT) score ( 64 ) or frequency of drinking ( 64 ), previous history of suicide attempt ( 82 ) or suicide among family members ( 60 ) or friends ( 82 ), lower family support or higher familial conflict ( 48 ), substance use in the previous two weeks ( 82 ), and demographic characteristics including age ( 64 , 65 , 83 ), lower educational level ( 65 , 82 , 83 ), and being female ( 82 , 83 ) were important when predicting suicide attempt.…”

Machine learning and the prediction of suicide in psychiatric populations: a systematic review

“… Xu et al, [ 41 ] 2323 patients with self-harm (1163/1160) and 46,460 inpatients controls (23,260/23,200) Self-harm (Not specified) Not specified Patient embedding method, Dx2Vec (Diagnoses to Vector) followed by NN 19 features from comorbidities Risk prediction at 12-month Precision 0.54 Sens: 0.72 for positive cases Dx2Vec-based model outperforms the baseline deep learning model in identifying patients who would self-harm within 12 months. Peis et al, [ 48 ] 1023 patients (662/361) Mixed diagnoses (Mood disorder 23%, anxiety disorders 53%) Not specified Recurrent NN 117 features from EMA in combination with traditional EHR Suicide ideation Acc: 95% AUC 0.94 Addition of EMA records boosts the prediction of suicidal ideation diagnosis from 48.13% obtained exclusively from EHR-based state-of-the-art methods to 67.78%. Gosnell et al, [ 81 ] 423 psychiatric inpatients (Not specified) Mixed diagnoses (mood, anxiety, personality, and SUD) Not specified RF (LOOCV) Structural (316) and resting-state functional connectivity (8256) measures Suicide attempts Sens: 79.4% Spec: 72.3% Altered resting-state functional connectivity features from frontal and middle temporal regions, as well as the amygdala, parahippocampus, putamen, and vermis were found to generalize best.…”

Universal mental health screening with a focus on suicidal behaviour using smartphones in a Mexican rural community: protocol for the SMART-SCREEN population-based survey

“…The main strength of this project is its innovative nature and its feasibility: the technology we will use in the SMART-SCREEN initiative has been previously used with success in clinical populations with mental disorders, achieving large sample sizes. [57][58][59][60] We are concerned that difficulties in access to SMART-SCREEN among people who do not have a smartphone, as well as among the elderly, may somewhat undermine our study. We hope to overcome this limitation with the involvement of local administrations and community leaders, who will disseminate the project and provide assisted computer stations in educational and municipal facilities.…”