From healthcare accessibility to health outcomes: A statistical and machine learning approach to large-scale graphs – GRAPH4HEALTH

The project is fourfold: it intends to (i) build a platform that can manage massive health data and make them usable for researchers; (ii) use the tools of graph theory to describe the healthcare system in a systematic and quantitative way; (iii) develop new machine learning tools to understand the shape of the graphs and predict their consequences on health outcomes; (iv) shed light on those policy issues that affect the efficiency of the French healthcare system.
We have access to all records of consultations and other medical procedures, drug prescriptions, and hospital admissions for the entire population living in France. The data cover the years 2008 to 2018. We will represent this unique data set as a series of time-evolving, geolocated, and bipartite graphs. Such graphs have two types of nodes: patients and a category of providers (e.g., generalist doctors). A patient and a provider are connected if they have met at least once during the current year. The projection of bipartite graphs on the set of providers informs about how patient sharing and referral networks.
We will develop econometric and machine learning methods to explain and/or predict the matching between patients and providers based on patients’ and providers’ characteristics (location, health condition, physician specialty, etc). Our two main goals are to understand the formation of the graphs and use these graphs to estimate the causal impact of the healthcare system on utilization and health outcomes (drug prescription, emergency hospital visits, mortality, etc.). We will examine whether certain local configurations are more effective at delivering better outcomes for patients. A particular attention will be paid to the geographic distribution of healthcare supply. We will build indicators of potential access at the local level, characterize potential low-density areas, the so-called “medical deserts”, and quantify their effects on patients’ outcomes.