Prediction model for suicide based on back propagation neural network and multilayer perceptron

Area of Science:

• Computational neuroscience
• Artificial intelligence in healthcare
• Psychiatric epidemiology

Background:

• Suicide remains a significant public health concern requiring accurate prediction models.
• Existing prediction methods often lack precision and invasiveness.
• Developing non-invasive, accurate suicide prediction tools is crucial for early intervention.

Purpose of the Study:

• To explore neural network models, specifically back propagation (BP) and multilayer perceptron, for suicide prediction.
• To establish a precise, non-invasive, and brief suicide prediction model.
• To compare the efficacy of different neural network architectures in suicide risk assessment.

Main Methods:

• Data collected via psychological autopsy (PA) in rural China.
• Univariate statistical methods used for preliminary factor selection.
• Back propagation neural network (BPNN) and multilayer perceptron models developed for prediction.

Main Results:

• The multilayer perceptron model achieved an AUC above 90%, outperforming the BPNN model (AUC ~82%).
• The multilayer perceptron demonstrated superior discrimination efficiency compared to BPNN.
• Neural network models showed higher accuracy than traditional logistic regression (80.9% correct).

Conclusions:

• Neural network prediction models offer greater accuracy than traditional methods for suicide risk assessment.
• The multilayer perceptron model is identified as the optimal prediction model for suicide.
• These findings support the clinical significance of neural network models for diagnosis and AI-assisted clinical systems.