Backpropagation and adaptive resonance theory in predicting suicidal risk

Area of Science:

• Psychiatry
• Computational Neuroscience
• Machine Learning in Healthcare

Background:

• Predicting suicide risk in major psychiatric patients is a critical challenge.
• Previous research has explored various statistical and machine learning models for suicide risk assessment.

Purpose of the Study:

• To investigate the efficacy of backpropagation and adaptive resonance theory (ART) neural networks in predicting suicide probability within a two-year timeframe.
• To evaluate the performance of these neural networks using historical patient data.

Main Methods:

• Collected data from 161 hospitalized psychiatric patients with a history of illness, including those with and without prior suicide attempts.
• Trained neural network systems using medically serious suicide attempts (MSSA) and non-MSSA classifications.
• Evaluated network performance by screening extreme cases: patients who committed suicide and those with no suicidal ideation, across three Israeli hospitals.

Main Results:

• Neither neural network system demonstrated reliable suicide prediction capabilities.
• Gehah Hospital records showed significantly better identification rates compared to two other hospitals (p < 0.05 for PPV; p < 0.01 for specificity).
• High reliability was observed when indicating low risk (NPV > 75%, specificity > 95%), but overall false positive rates were significant.

Conclusions:

• Current neural network models are not sufficiently reliable for evaluating suicidal risk due to a high number of false positives.
• Data from Gehah Hospital suggests that incorporating direct-subjective questionnaires might improve future prediction accuracy.
• Both ART and backpropagation models performed similarly across all evaluated metrics.