Learning Hidden Patterns from Patient Multivariate Time Series Data Using Convolutional Neural Networks: A Case Study of Healthcare Cost Prediction

Objective: To develop an effective and scalable individual-level patient cost\nprediction method by automatically learning hidden temporal patterns from\nmultivariate time series data in patient insurance claims using a convolutional\nneural network (CNN) architecture.\n Methods: We used three years of medical and pharmacy claims data from 2013 to\n2016 from a healthcare insurer, where data from the first two years were used\nto build the model to predict costs in the third year. The data consisted of\nthe multivariate time series of cost, visit and medical features that were\nshaped as images of patients' health status (i.e., matrices with time windows\non one dimension and the medical, visit and cost features on the other\ndimension). Patients' multivariate time series images were given to a CNN\nmethod with a proposed architecture. After hyper-parameter tuning, the proposed\narchitecture consisted of three building blocks of convolution and pooling\nlayers with an LReLU activation function and a customized kernel size at each\nlayer for healthcare data. The proposed CNN learned temporal patterns became\ninputs to a fully connected layer.\n Conclusions: Feature learning through the proposed CNN configuration\nsignificantly improved individual-level healthcare cost prediction. The\nproposed CNN was able to outperform temporal pattern detection methods that\nlook for a pre-defined set of pattern shapes, since it is capable of extracting\na variable number of patterns with various shapes. Temporal patterns learned\nfrom medical, visit and cost data made significant contributions to the\nprediction performance. Hyper-parameter tuning showed that considering\nthree-month data patterns has the highest prediction accuracy. Our results\nshowed that patients' images extracted from multivariate time series data are\ndifferent from regular images, and hence require unique designs of CNN\narchitectures.\n