Back

dynaPhenoM: Dynamic Phenotype Modeling from Longitudinal Patient Records Using Machine Learning

Zhang, H.; Zang, C.; Xu, J.; Zhang, H.; Fouladvand, S.; Havaldar, S.; Su, C.; Cheng, F.; Glicksberg, B. S.; Chen, J.; Bian, J.; Wang, F.

2021-11-02 health informatics
10.1101/2021.11.01.21265725 medRxiv
Show abstract

Identification of clinically meaningful subphenotypes of disease progression can facilitate better understanding of disease heterogeneity and underlying pathophysiology. We propose a machine learning algorithm, termed dynaPhenoM, to achieve this goal based on longitudinal patient records such as electronic health records (EHR) or insurance claims. Specifically, dynaPhenoM first learns a set of coherent clinical topics from the events across different patient visits within the records along with the topic transition probability matrix, and then employs the time-aware latent class analysis (T-LCA) procedure to characterize each subphenotype as the evolution of these learned topics over time. The patients in the same subphenotype have similar such topic evolution patterns. We demonstrate the effectiveness and robustness of dynaPhenoM on the case of mild cognitive impairment (MCI) to Alzheimers disease (AD) progression on three patient cohorts, and five informative subphenotypes were identified which suggest the different clinical trajectories for disease progression from MCI to AD.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.