Topic Models for Primary Care and Community Health Monitoring

Topic models are a class of unsupervised machine learning algorithm used to estimate latent thematic patterns underlying large unstructured document collections. This thesis investigates how topic models can leverage growing amounts of routinely collected clinical text data for descriptive characterization and temporal monitoring of primary care and community health. To begin, the thesis demonstrates how non-negative matrix factorization (NMF) can be used to estimate a meaningful thematic summary of the primary care system (e.g. in terms of patients archetypes and characteristic activities). The NMF estimated latent parameter matrices are transformed into a multivariate time series data structure and AR-1 dynamic regression models are used to monitor the temporal evolution of primary care topical vectors. The unique text-based descriptive time series design identified diverse COVID-19 pandemic effects on primary care and community health in Toronto, Canada (many of which have been corroborated by independent research groups using alternative study designs, data sources, and statistical methods). Next, the thesis transitions to evaluate topic model validity. In our context, validity refers to the extent to which empirical evidence and existing theory support the interpretations of estimated latent topical vectors for their purported use cases (i.e. descriptive characterization and temporal monitoring of primary care and community health). One study investigates concurrent validity by correlating estimated primary care thematic vectors against contemporaneously collected most responsible diagnostic codes. Another study examines temporal/geographic stability of latent topical vectors estimated using two distinct primary care clinical text corpora from Ontario, Canada. A final study explores whether inferences regarding characterization and monitoring of primary care thematic content are robust to the choice of statistical topic model and associated framework for estimation/inference. The collection of interrelated research studies included in the thesis illustrate how topic models and clinical text data can be used to generate meaningful insights regarding characterization and temporal monitoring of primary care andcommunity health.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.