Introduction Rapid advances in computing power and machine learning techniques, coupled with the widespread adoption of connected consumer devices and sensors (smartphones, smartwatches) and health information systems (electronic health records, personal health records), have created ideal conditions for the emergence of new digital diagnostic algorithms. These are pattern-searching programs that use routinely collected health data to predict the appearance of future diseases or to detect early manifestations of subclinical diseases. Such algorithms are being applied in an increasing range of cases, including the use of smartwatches to screen for atrial fibrillation, the use of typing patterns on a keyboard to forecast Parkinson’s Disease, and the analysis of electronic health record data to predict which patients will develop sepsis in the hospital. They can be developed and deployed with relative ease, both by medical specialists and data scientists. While on the one hand, such digital tools may help to better detect and prevent real disease before it harms patients or becomes less treatable, on the other hand, these tools will increasingly be applied to healthy individuals to detect conditions that may never have harmed them, leading to an increase in digital overdiagnosis.

Structure In this talk, we will a) describe the concept of digital overdiagnosis as an emerging phenomenon, b) highlight the lack of awareness and consensus in the medical and data analytics communities and the need for greater dialogue between them, c) underline the need to provide a data-centered definition of overdiagnosis, and d) propose strategies to quantify the risk of digital overdiagnosis as the first essential step to address and mitigate it during development of digital diagnosis tools.

Expected Results We aim to raise awareness of the nature and risk of an important new form of overdiagnosis—digital overdiagnosis—and to establish a preliminary research agenda to address the problem. This research agenda involves exploring how data and machine learning can itself be used to mitigate the overdiagnosis risks caused by the explosion in clinical data availability and diagnostic algorithms. The benefits of preventing digital overdiagnosis include helping reduce a precipitous rise in public health expenditure and limiting individual patient and population exposure to the risks of over-investigation and overtreatment, which are likely to be increasingly generated by this techno-medical phenomenon.