DeepDeath: Learning to predict the underlying cause of death with Big Data

Area of Science:

• Public Health
• Data Science
• Computational Biology

Background:

• Multiple cause-of-death data is crucial for public health surveillance and predicting population health trends.
• Large datasets of mortality data are available but require advanced Big Data techniques for analysis.
• Existing methods may not fully capture the complex temporal dynamics within mortality data.

Purpose of the Study:

• To develop and evaluate Big Data models for large-scale analysis of U.S. mortality data.
• To compare the performance of a Hadoop-based ensemble of random forests with a deep learning model.
• To assess the capability of models in uncovering hidden patterns and temporal aspects of mortality data.

Main Methods:

• Designed two classes of models: a Hadoop-based ensemble of random forests trained over N-grams and DeepDeath, a deep classifier using recurrent neural networks (RNNs).
• Applied both models to mortality data from the National Center for Health Statistics.
• Utilized long short-term memory (LSTM) networks within the DeepDeath model to capture temporal dependencies.

Main Results:

• Both developed models significantly outperformed a random classifier in analyzing mortality data.
• The DeepDeath model, employing LSTMs, demonstrated superior performance compared to the N-gram based random forest models.
• The deep learning approach effectively learned the temporal characteristics of the data without requiring manual feature engineering.

Conclusions:

• Deep learning models, particularly those using LSTMs, are highly effective for analyzing large-scale mortality datasets.
• Advanced computational techniques can unlock deeper insights from public health data, improving health standard predictions.
• The DeepDeath model offers a powerful, data-driven approach to understanding mortality trends and temporal patterns.