Deep Sequential Models for Suicidal Ideation From Multiple Source Data.

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Digital phenotyping

Background:

• Predicting suicidal ideation is crucial for mental healthcare.
• Electronic Health Records (EHR) and Ecological Momentary Assessment (EMA) offer rich, longitudinal patient data.
• Existing methods struggle with the asynchronous and variable nature of EHR and EMA data.

Purpose of the Study:

• To develop a novel deep sequential model for predicting suicidal ideation.
• To integrate both EHR and EMA data for enhanced prediction accuracy.
• To provide interpretability for the prediction model.

Main Methods:

• Utilized recurrent neural networks (RNNs) to model longitudinal EHR and EMA data sequences.
• Developed a temporal alignment strategy by concatenating RNN hidden states.
• Incorporated attention mechanisms for long sequences and pre-trained embeddings for short sequences.
• Validated the model on a database of 1023 patients.

Main Results:

• The proposed method significantly improved the recall for predicting suicidal ideation.
• System recall increased from 48.13% (EHR-only) to 67.78% with the addition of EMA data.
• The model demonstrated interpretability via t-distributed stochastic neighbor embedding (t-SNE) visualization.
• Key input features were identified and medically interpreted.

Conclusions:

• Combining EHR and EMA data with deep sequential models substantially enhances suicidal ideation prediction.
• The developed method offers a promising, interpretable approach for clinical risk assessment.
• Future work can explore further integration of multimodal data for improved mental health predictions.