Abstract: Surveillance is still a productive topic in public health informatics but other very important topics in Public Health are appearing. For example, the use of artificial intelligence approaches is increasing.
Introduction
Public health is undergoing a significant paradigm shift, moving beyond traditional, structured surveys toward the era of "digital epidemiology." This evolution leverages massive, unstructured datasets—including social media streams, internet search trends, and electronic health records (EHRs)—to monitor population health in real-time. While these "Big Data" sources offer a pulse on the population that traditional clinical reports often miss, they present two primary hurdles: the difficulty of extracting actionable insights from noisy data and the critical need to protect patient privacy when utilizing clinical notes.
Methodological Landscape
The 2018 IMIA Yearbook section on Public Health and Epidemiology Informatics reviewed the state of the art by analyzing 843 references from 2017. The review highlights a transition toward sophisticated computational techniques designed to handle the volume and variety of modern health data. Key methodologies identified include:
- Natural Language Processing (NLP): Researchers are utilizing Latent Dirichlet Allocation (LDA) for topic modeling and Recurrent Neural Networks (RNNs)—specifically Long Short-Term Memory (LSTM) networks—to automate the de-identification of sensitive information in clinical text.
- Causal Inference: Techniques such as Transfer Entropy are being applied to determine the direction of influence between media reporting and public emotional responses.
- Digital Surveillance Integration: Open-source platforms (like Ushahidi) are being paired with "geotracers" and SMS-based reporting to track non-infectious events, such as road traffic crashes, alongside traditional infectious disease monitoring.
Key Insights and Case Studies
The Media-Fear Feedback Loop
One of the most significant findings highlighted in the review (Choi et al.) involved the 2015 MERS outbreak in Korea. By analyzing 153 news outlets, researchers discovered a positive feedback loop: media outlets overestimated the lethality of the virus, which triggered public fear; this fear, in turn, motivated further sensationalist reporting. This cycle caused substantial social and economic disruption, demonstrating how AI can help quantify the impact of information flow on public behavior.
AI vs. Traditional Statistics
In the realm of data privacy, deep learning has proven superior to traditional methods. When redacting patient names and addresses from EHRs (using the i2b2 challenge dataset), RNNs significantly outperformed Support Vector Machines (SVMs) and Conditional Random Fields (CRFs). The neural networks' ability to "understand" linguistic context makes them a more flexible and accurate tool for solving the data-sharing bottlenecks that often hinder epidemiological research.
Predictive Capabilities
AI is also being used to anticipate public health crises before they reach a breaking point. By modeling the interaction between vaccine decisions and disease dynamics as a nonlinear feedback loop, researchers have used Google and Twitter data to identify "critical transition" points where public resistance to vaccination might lead to an outbreak. Furthermore, integrating NASA climate data with clinical records has allowed for more accurate modeling of seasonal variations in diseases like malaria and influenza.
Limitations and Future Directions
Despite these advancements, the field faces several challenges that the next generation of researchers must address:
- Generalizability: Many AI models are trained on specific outbreaks (like MERS) or within specific geographic regions. Their accuracy across different cultural or linguistic contexts remains largely unproven.
- Impact Measurement: There is a critical need for indicators that measure whether AI-driven insights actually lead to improved public health performance or tangible policy changes.
- Methodological Rigor: The authors emphasize that the field must move away from the simple collection of massive web-generated data and return to asking "clever and appropriate questions" paired with rigorous validation against traditional gold standards.
Conclusion
Artificial intelligence is rapidly becoming an essential component of the public health toolkit. By automating de-identification and providing real-time sentiment and outbreak analysis, AI allows health officials to respond more quickly to emerging threats. However, the future of the field depends on ensuring these models are generalizable, validated, and—most importantly—capable of driving meaningful public health interventions.
Reference: Thiébaut R, Thiessard F. Artificial Intelligence in Public Health and Epidemiology. IMIA Yearbook of Medical Informatics 2018: 207-210.
Summary updated