Federated Artificial Intelligence In Big Data And Healthcare Using Real-World Setting

Multiple sclerosis (MS) is an inflammatory disorder of the brain and spinal cord, where the body's immune system incorrectly attacks its own central nervous system (CNS), causing variable and unpredictable symptoms. Some examples of clinical symptoms include slurred speech, blurred vision, loss of balance, poor coordination, tremors, numbness, extreme fatigue, problems with memory and concentration, paralysis, and blindness. MS affects more than 700,000 people living in Europe, two to three times as many women as men, and is most often diagnosed in people aged 20-40 years. MS varies in severity, from mild symptoms to a disabling condition. The overall cost of MS in Europe to health and social care is estimated to be 15 billion euros per year (Kobelt, Thompson et al. 2017). The management of MS disease comprises a wide range of drugs with different modes of action, as well as a broad spectrum of proposed rehabilitation strategies, resulting in varying levels of efficacy that need meticulous monitoring. Today, disease management of people with MS (PwMS) is based on insights gained from population-based studies. To date, 14 DMTs have been approved for relapsing-remitting MS on the basis of their efficacy in randomized controlled trials (RCTs). RCTs are accepted as the gold standard for assessing the efficacy and safety of any new drug and are conducted in a controlled setting with well-defined homogeneous patient populations selected through strict inclusion criteria. These cohorts do not necessarily represent MS in real life and conclusions made from these RCTs therefore do not always translate to the individual patient. High-performance MS-specific decision support systems are needed to support treatment decision-making by neurologists and regulators (= the right DMT for the right patient). The time needed to find the matching DMT for a given PwMS is still mostly based on trial and error in clinical practice. Moreover, a treatment that was initially satisfactory might eventually become ineffective because of disease progression. Delays in identifying treatment failure and in selecting an appropriate next-line DMT will further degrade the quality of life of patients. Real-world data (RWD) are defined as data derived from a number of sources that are associated with outcomes in a heterogeneous patient population representing the real-world settings (e.g., data collected by physicians in standard clinical care and longitudinal follow-up) (Trojano, Tintore et al. 2017). Despite the increasing reliance on RWD, challenges and limitations exist that complicate the generation, collection, and use of this data. First, data points are usually not evenly spaced in time because medical appointments are not scheduled at a constant rate. This breaks down a typical assumption used in many common modelling techniques. Second, RWD observations typically contain a large fraction of unmeasured data. Indeed, all variables are not measured at each (clinical) appointment, leading to scarcely observed time series (e.g., typically magnetic resonance imaging (MRI) are measured once every year, whereas some neurologists measure electrophysiological responses every 3 months). Moreover, the observation pattern is expected to be informative by itself as it directly reflects the neurologist’s medical practice (e.g., more frequent measurement of MRI might reflect a neurologist’s concern for the appearance of new brain lesions and thus be associated with a higher probability of new brain lesions). Third, the data usually consists of a combination of events and continuous measurement (e.g., ‘relapse’ versus ‘disability scores’). Lastly, as we consider data from less controlled environments, measurements are more prone to noise and to incorrect coding. Community efforts are undertaken to develop decision support systems for DMT effectiveness at the patient level, but the existing tools or algorithms either do not include the effect of DMTs (Daumer, Neuhaus et al. 2007, Galea, Lederer et al. 2013, Signori, Izquierdo et al. 2017) or they only predict short-term responses (e.g., 6 months) with limited precision and large confidence intervals (Kalincik, Manouchehrinia et al. 2017). Most recent works manually extract temporal features, such as minimum expanded disability severity score (EDSS) or number of relapses in the last year, without ever considering the fundamental difference that arises when observing the full patient trajectory. Yet, medical progression in MS is known to be of utmost relevance for clinical decisions (Signori, Izquierdo et al. 2017). Recent work performed by our group shows that introducing machine learning methods that take into account the full patient trajectory improves predictions of disability progression in MS. Using full medical trajectories of over 9,000 patients with MS, we were able predict more accurately the disease progression than previous state of the art attempts. By modelling full patient history rather than temporal summary statistics, our method reached an AUC of 0.85 for patient degradation over the next 2 years with a 3 years history observation window. However, to develop algorithms that are accurate and precise enough to reach insights at the individual level, an extensive amount of data is required. The use of a combination of very large and fit-for-purpose existing cohorts and/or registry could overcome some of the statistical limitations of RWD (=’scale-up’). Data sharing and scaling-up health data research is not easy. Policies of data sharing should rest upon knowledge of how data is shared and how end-users use data that have been shared to them. Here, we propose to introduce a federated artificial intelligence approach to combine data sets. Federated artificial intelligence denotes the distribution of the learning effort over physically separated partners. This goes beyond the currently more established concept of “federated databases” where the data are distributed, but not the learning from the data. It is key to enable custodian control over data during learning. In this project, the importance of federated artificial intelligence for MS patient trajectories will be demonstrated.