JMIR Publications

ABSTRACT

Background:

Health care systems are amassing increasing volumes of digital text information on their patients. Topic models are a class of statistical model which can assist with organization, browsing and characterization of the latent patterns permeating these large clinical text collections.

Objective:

To evaluate several methods for estimating temporal topic models, using clinical notes obtained from primary care electronic medical records (EMRs).

Methods:

We used a retrospective closed cohort design. The study timeframe spanned 01-Jan-2011 through 31-Dec-2015, discretized into 20 quarterly periods. Patients were included in the study if they generated at least one primary care clinical note in each of the 20 quarterly time-periods. These patients represented a unique cohort of individuals engaging in high frequency use of the primary care system. The following temporal topic modelling algorithms were fitted to the clinical note corpus: non-negative matrix factorization (NMF), latent Dirichlet allocation (LDA), the structural topic model (STM), and the BERTopic model.

Results:

Temporal topic models consistently identified latent topical patterns in the clinical note corpus. The learned topical bases identified meaningful activities conducted by the primary healthcare system. Latent topics displaying near-constant temporal dynamics were consistently estimated across models (e.g. pain, hypertension, diabetes, sleep/mood/anxiety/depression). Several topics displayed predictable seasonal patterns over the study timeframe (e.g. respiratory disease, influenza immunization programs).

Conclusions:

NMF, LDA, STM, and BERTopic are based on different underlying statistical frameworks (e.g. linear algebra and optimization, Bayesian graphical models, and neural embeddings), require tuning unique hyper-parameters (e.g. optimizers, priors, etc.) and have distinct computational requirements. Despite the heterogeneity in statistical methodology, the learned latent topical summarizations, and their temporal evolution over the study timeframe, were consistently estimated. Temporal topic models represent an interesting class of models for characterizing and monitoring the primary healthcare system.