Machine learning to refine decision making within a syndromic surveillance service.

Area of Science:

• Public Health
• Epidemiology
• Machine Learning

Background:

• Syndromic surveillance systems generate numerous statistical alarms daily for public health threat detection.
• Current risk assessment processes are time-consuming and require specialized expertise, with only 0.1% of alarms being of public health importance.

Purpose of the Study:

• To evaluate machine learning as a tool for computer-assisted decision-making in assessing statistical alarms within syndromic surveillance.
• To determine if machine learning can enhance the automation, robustness, and rigor of risk assessment processes.

Main Methods:

• Utilized a dataset of 67,505 statistical alarms and their corresponding risk assessment outcomes from Public Health England (August 2013 - October 2015).
• Employed three Bayesian classifiers (naïve Bayes, tree-augmented naïve Bayes, Multinets) to analyze alarm characteristics and predict epidemiologist decisions ('Alert', 'Monitor', 'No-action').
• Implemented additional classifications to address the high proportion of 'No-action' outcomes.

Main Results:

• The naïve Bayes classifier achieved the highest performance, correctly classifying 51.5% of 'Alert' outcomes.
• Combining 'Alert' and 'Monitor' classifications improved performance to 82.6% accuracy.
• Demonstrated the potential for operationalizing a decision support system using a naïve Bayes classifier within syndromic surveillance.

Conclusions:

• Machine learning techniques offer significant potential to automate and improve the risk assessment of statistical alarms in syndromic surveillance.
• Specialist human input remains essential for the effective interpretation and validation of public health alarm assessments.