Abstract:
|
Explaining subject-specific outcome is important for personalized or precision medicine. Linear mixed effects models (LMM) are often effective for identifying and incorporating subject-specific randomness from its general population. However, traditional methods to fit a LMM require the entire data to be used at once. With electronic health records (EHR), we face challenges in using LMM because the EHR are often huge and can exceed memory capacity allowed in standard statistical programs. Our bLMM is a method for fitting linear mixed effects model for big longitudinal data. bLMM finds estimates using partitioned data and by sequentially updating the estimates between and within the partitions using a partial EM algorithm. bLMM works for a general LMM with a random intercept, multiple random slopes and/or higher-order random terms, which can be correlated. Also, bLMM allows an additional serial correlation per subject for repeated measures. For implementation, some preliminary studies of bLMM will be provided. bLMM is a powerful way to explain subject-specific outcomes from a big data.
|