Development of a Machine Learning Model Using Multiple, Heterogeneous Data Sources to Estimate Weekly US Suicide...

Area of Science:

• Public Health
• Data Science
• Epidemiology

Background:

• Suicide is a major public health concern in the US.
• Official suicide statistics are delayed by 1-2 years, hindering timely public health interventions.
• There is a critical need for near real-time data on suicide fatalities.

Purpose of the Study:

• To develop and validate a machine learning pipeline for estimating weekly suicide fatalities in the US in near real time.
• To improve the timeliness of suicide data for public health planning and decision-making.

Main Methods:

• A cross-sectional national study utilizing a two-phase machine learning approach.
• Phase 1: Fitting optimal machine learning models to individual data streams (e.g., syndromic surveillance, lifeline calls, economic data, online trends).
• Phase 2: Combining stream predictions using an artificial neural network to estimate weekly suicide fatalities.

Main Results:

• The machine learning pipeline demonstrated a high correlation (Pearson correlation, 0.811) with actual weekly suicide counts and trends.
• The model estimated annual suicide rates with a low error of 0.55%.
• The ensemble framework significantly reduced estimation error compared to traditional forecasting methods.

Conclusions:

• The developed machine learning framework provides accurate, near real-time estimates of suicide fatalities.
• This novel approach offers potential for more effective public health responses, including resource allocation and intervention deployment.
• The findings highlight the utility of integrating diverse data streams for critical public health surveillance.