Compressive Population Health: Cost-Effective Profiling of Prevalence for Multiple Non-Communicable Diseases via Health Data Science

With a growing ageing population and changes in lifestyles, non-communicable diseases (NCD), e.g. heart disease, diabetes, and cancer, have become extremely prevalent in our society, and the situation is more challenging in UK compared to other developed countries. Population health monitoring is fundamental block for public health services, and profiling population-scale prevalence of multiple NCD across different regions (e.g., building the spatially fine-grained morbidity rate map) is one of the most important tasks. However, traditional public health data collection and prevalence profiling approaches, such as clinic-visit-based data integration and health surveys, are often very costly and time-consuming. This project proposes a novel paradigm, called compressive population health (CPH for short), to reduce the data collection cost during the profiling of prevalence to the maximum extent.

The basic idea CPH is that a subset of areas is intelligently selected for data collection and population health profiling in the traditional way, while leveraging inherent data correlations to perform data inference for the rest of the areas. CPH is facilitated by the exploitation of the following types of inherent data correlations found by epidemiologists. (a) Intra-Disease Spatial Correlations. That is, regions are more similar in the prevalence rate of some diseases when they are neighbouring, or share certain common environmental, socioeconomic, and demographical attributes. (b) Inter-Disease Correlations. Multimorbidity, commonly defined as the co-presence of two or more chronic conditions, demonstrates that statistics for different types of disease may also correlate with each other. For example, regions with higher obesity rate are more likely to have higher rates of heart disease and cancers.

In order to realize this idea, this project develops three technical work packages to accomplish the following technical goals: (1) Investigate and extract latent data correlations and further utilize them to build learning models for prevalence inference on the target geographical grids. (2) Design intelligent algorithms for selecting traditional-sensed areas for each disease with multi-objective optimization goals including cost, reliability, and latency. (3) Evaluate and interpret the inference results of prevalence rate to ensure the reliability and robustness of the approach.

The proposed CPH is a novel solution to a public health data collection challenge enabled by data science and artificial intelligence. It opens the door for a disruptive population health monitoring paradigm with potential significant cost reductions for public health authorities. By closely working with partners from public health sector, including NHS England and Public Health at Warwickshire County Council, we will evaluate the feasibility of this approach based on multiple public health datasets together with relevant demographic/geographic statistics in the same regions.