A Landmark Competing Risk Model for Dynamic Prediction

Oregon Convention Center 

Electronic health records (EHRs) are promising but challenging resources for research on investigating and monitoring disease progression. Motivated by the hospitalized COVID-19 patient data from West Bengal in India, we aim to dynamically predict the chance of "discharge" or "death" of these COVID-19 hospitalized patients based on their longitudinal laboratory measurements. In total, there are 147,805 hospitalized COVID-19 patients with 1,091,322 laboratory measurements, and the high volume of this data raises the computation challenge for dynamic prediction. In addition, the features of EHRs data such as sparsity, irregularity and non-linearity also place a challenge in modelling. To address these, we propose a two-step landmark competing risk model which summarizes the historical laboratory measurements using a functional principle analysis (PCA) and then uses the landmark competing risk model for prediction. The proposed method is easy to implement using the existing software. All estimated model parameters, longitudinal history, and at-risk population vary over the landmark time. The whole dataset was randomly split into training and testing set with the ratio of 1:1. Different approaches for handling longitudinal observations including baseline measure, mean, recent measure (last value carry forward), and linear regression are adopted in the two stage estimation and compare with the proposed method via the weighted Harrell's C-Index and Brier score. The proposed method outperforms all comparable methods at the distant landmark time. Using the proposed model we dynamically predict "death" or "discharge" given the different landmark time and depict their associations with COVID-19 medication according to their historical laboratory measurements, which provide the evidence that this model has potential to assist clinicians in understanding patients' disease progression at different time and providing the suggestion about the medication use based on their historical information.

landmark model

competing risk

dynamic prediction

longitudinal data

survival 

Section on Nonparametric Statistics