Using Neural Networks with Routine Health Records to Identify Suicide Risk: Feasibility Study.

Area of Science:

• Medical Informatics
• Public Health
• Artificial Intelligence

Background:

• Suicide remains a significant global health issue, with 800,000 deaths annually.
• Identifying individuals at high risk for suicide is challenging due to complex factors.
• Many individuals who die by suicide interact with health services prior to their death.

Purpose of the Study:

• To assess the feasibility of using artificial neural networks (ANNs) with electronic health records (EHRs) for suicide risk identification.
• To support early identification of high-risk individuals within healthcare systems.

Main Methods:

• Utilized UK Secure Anonymised Information Linkage Databank for suicide cases (2001-2015) and controls.
• Extracted data from primary and secondary EHRs, creating feature vectors of risk factors.
• Trained ANNs to differentiate suicide cases from controls, interpreting output as suicide risk score.

Main Results:

• The best ANN system achieved a 26.78% error rate, with 64.57% sensitivity and 81.86% specificity.
• Prescription of psychotropics, depression, anxiety, and self-harm significantly increased estimated suicide risk.
• These factors identified at least 95% of suicide cases, indicating strong predictive value.

Conclusions:

• The ANN methodology demonstrated comparable accuracy to existing methods using specialized data.
• Heterogeneity in suicide case patterns contributed to errors, with some cases resembling controls.
• Key risk factors include psychotropic prescriptions, mental health diagnoses, self-harm, hospital admissions, and substance misuse.