Anomaly Detection in Endemic Disease Surveillance Data Using Machine Learning Techniques.

Area of Science:

• Epidemiology
• Machine Learning
• Public Health Surveillance

Background:

• Disease surveillance data is underutilized for real-time decision-making.
• Early detection of outbreaks and intervention priorities are crucial for disease control.

Purpose of the Study:

• To explore unsupervised anomaly detection machine learning techniques for discovering epidemiological signals.
• To assess the potential of these methods in improving disease surveillance, using the Brazilian Amazon malaria dataset as a case study.

Main Methods:

• Applied unsupervised anomaly detection machine learning models to the Brazilian Amazon malaria surveillance dataset.
• Evaluated the models' ability to detect outbreak onset, peaks, and changes in positive case proportions.
• Determined the optimal number of models (top-k) to maximize anomaly detection across different health regions.

Main Results:

• Machine learning models successfully provided early indications of malaria outbreak onset, peaks, and change points in 2016.
• No single model detected all anomalies across all health regions, highlighting the need for ensemble approaches.
• Principal Component Analysis, Stochastic Outlier Selection, and Minimum Covariance Determinant were identified as the top three models for maximizing anomaly detection coverage.

Conclusions:

• Anomaly detection is a valuable approach for identifying epidemiological patterns in large spatio-temporal datasets.
• The proposed methodology can be replicated for other diseases and locations to enhance monitoring and timely interventions.
• This approach supports efforts towards controlling endemic diseases through improved surveillance and data utilization.