203 Summary and general discussion 9 known example of an innovative surveillance system that turned out to be unreliable, which was caused by overfitting4,16. Google Flu Trends missed large outbreaks and wrongly predicted non-existing ones, potentially because people’s changing online search behaviour was not accounted for4,18. An intriguing question is how Artificial Intelligence (AI) is going to change disease surveillance in the (near) future. AI has been around for decades19, but has become better, cheaper, and more easily accessible in recent years20. Even with automated data collection and aggregation as described above, there are still labour-intensive steps that require humans, like data interpretation. Given the exponential growth of cases in an epidemic, early intervention is essential, and AI could potentially improve early detection21. Data sources that are impossible for humans to analyse because of the size, are easy for computers. With the developments in natural language processing, the collection, filtering and analyses of data from various (open) online sources (e.g., social media, news outlets, Wikipedia page views) and hospital data (e.g., discharge letters) creates many new opportunities18,21–23. Apart from these (unstructured) data sources, AI can also interpret images and videos, which could add to traditional (structured) data that is typically used for surveillance23. This could allow for earlier detection of outbreaks, detection of regional patterns in outbreaks, and for sophisticated surveillance systems in countries or areas with fewer resources21. HealthMap is an example of an AI algorithm, that scans various online sources (such as search queries, social media, news reports) for signs of disease outbreaks18. On December 30, 2019, a few days after the first COVID-19 case was identified in Wuhan, China, it provided the first alarm for a potential outbreak, before any humans did18,22. The idea to utilise AI, specifically machine learning methods, to detect patterns and analyse trends is not new, but may be increasingly needed to analyse complex data23. AI could potentially add to traditional infectious disease surveillance by generating early detection warnings, assist with case finding by mining data sources that are difficult for humans to analyse, and interpret complicated geographical information or data sources that generate a constant data stream, like wearable devices that monitor heart rate, number of steps or oxygen saturation22,23. Another aspect in which AI could advance infectious disease surveillance is through the improvement in epidemic modelling and simulations23. However, there are also several important concerns regarding the use of AI. These concerns are not limited to medicine or infectious diseases, but are relevant to consider nonetheless. First of all, the principle of “garbage in- garbage out” also holds true in the context of AI, as illustrated by the previously mentioned examples from Google Flu Trends. A second important concern is regarding the introduction of bias20. There are many examples of AI models being trained on non-representative data, which can lead to issues of racism and sexism25. For instance, when models are trained to identify skin infections using sample photos featuring mostly lighter skin tones, their accuracy to recognise these skin infections on darker skin tones may be compromised. Moreover, models trained on data from first-world countries, may not be applicable in third-world countries, potentially increasing existing health inequalities25. It is unlikely that AI can completely replace traditional infectious disease
RkJQdWJsaXNoZXIy MTk4NDMw