Today's NEJM contains a detailed review article on applications of AI in infectious disease surveillance. The Boston-based group is reported as Brownstein et al. NEJM 388:1597. It's part of a new series of topics for AI in medicine.
The articles discuss the role of AI in infectious-disease surveillance, with examples from Covid-19 tracking. AI methods, such as reinforcement learning, can optimize testing resources and aid early detection. However, overfitting, underrepresentation of certain populations, and privacy concerns present challenges. The future of surveillance involves emerging technologies like large language models, yet success relies on understanding the limitations of AI, ensuring representativeness in datasets, and prioritizing international cooperation.
GPT4 / Long
Part One discusses the application of artificial intelligence (AI) in infectious disease surveillance, particularly in the context of the Covid-19 pandemic. It emphasizes the use of AI and machine learning in analyzing vast amounts of data in real-time, which is beyond human capacity. This includes data derived from various sources such as social media, mobile devices, internet search trends, electronic health records, and genomics databases. These data sources can be classified as syndromic surveillance, digital surveillance, and genomic surveillance.
Syndromic surveillance involves collecting health-related data before official diagnoses are made, which can provide early indications of an outbreak. Digital surveillance uses data from sources like social media and internet searches, tracking disease-related search terms or symptom mentions to detect and monitor outbreaks. Genomic surveillance is based on sequencing pathogen genomes, allowing for tracking of disease spread, identification of new variants, and informing vaccine design.
AI-enhanced surveillance can help identify unusual disease patterns and predict future trends. For instance, during the Covid-19 pandemic, AI models predicted the trend of the disease, identified disease hotspots, and facilitated public health responses. The Part One also emphasizes the importance of developing AI models that can adapt to new challenges and identify signals amidst noise.
Part Two focuses on the risk assessment of infectious diseases, and the application of AI in managing and reducing disease transmission. It provides examples of how AI has been used in real-world scenarios, like health QR codes in China for real-time risk assessment and AI-driven chatbots for health-related queries. It also describes the use of AI in Greece for border surveillance, where an AI algorithm called Eva was used to screen travelers for Covid-19, outperforming tests based on epidemiologic metrics.
The application of reinforcement learning in Eva is discussed, where the algorithm sorts travelers according to various factors like origin country, age, sex, and time of entry. Test results are used to update the system and assign travelers for Covid-19 testing based on recent prevalence estimates for their type. The system continuously learns and optimizes the allocation of limited testing resources.
The text then moves onto exploring extended applications, including wearable technology for early infection detection, pinpointing outbreak hotspots, and leveraging AI for passive surveillance of adherence to non-pharmaceutical interventions. These AI-driven approaches can complement traditional public health surveillance, which is accurate but slower, and participatory surveillance, which provides real-time insights but lacks the confirmatory nature of traditional reporting.
However, the document also acknowledges several challenges and roadblocks for AI in surveillance. Data volume and quality are crucial for surveillance, but there can be issues with overfitting (tailoring predictions too specifically to existing data) and difficulties in distinguishing similar clinical manifestations without molecular testing. Surveillance systems need to be recalibrated regularly to adapt to new pathogen variants and changing external factors like vaccination.
There is also a concern about underrepresentation of certain populations in databases, which can lead to skewed and inequitable outcomes. This is both a moral and a methodological issue that needs addressing. The document provides an example of an AI trained on a database of chest radiographs in children which ended up distinguishing adults from children rather than diagnosing Covid-19.
Privacy issues arise as surveillance models incorporate more data streams from digital interactions and connected devices. The document suggests the use of federated learning, a technology that allows for local calculations on personal data without centrally storing it.
The document ends by emphasizing that AI cannot replace the cross-jurisdictional and cross-functional coordination required for combating emerging diseases, despite its power in improving surveillance infrastructure. The future of infectious-disease surveillance will rely on emerging technologies like biosensors, quantum computing, and augmented intelligence, including large language models like GPT-4. However, the success of these next-generation AI-driven surveillance tools will depend heavily on our ability to unravel their shortcomings and identify which achievements can be generalized.