Supervised multi-specialist topic model with applications on large-scale electronic health record data

MixEHR-Guided: A guided multi-modal topic modeling approach for large-scale automatic phenotyping using the electronic health record

“…Moreover, while supervised topic models such as MixEHR-S [61] In proof-of-concept analyses, we observe that the MixEHR-G-inferred topics are wellaligned with the phenotypes they represent and complementary to rule-based phenotyping algorithms. We used MixEHR-G trained on the PopHR database to derive meaningful insights about disease similarities, comorbidities, and prevalences in the Quebec population [62].…”

“…While the latent factors or topics inferred by these unsupervised methods can provide clinical insights, they are not identifiable as inferred topics cannot be directly mapped to known phenotypes. Moreover, while supervised topic models such as MixEHR-S [61] can simultaneously infer topic distributions and fit a predictive function of a target disease, they are not scalable to predicting multiple disease labels simultaneously. Therefore, an efficient method is needed to achieve simultaneous phenotype inference while providing interpretable topic distributions over heterogeneous EHR data.…”

“…For patient outcome classification, the inferred phenotype topic probabilities can be treated as phenotype risk scores or PheRS [56] derived from the training patient dataset. As a post-processing step, they can then be incorporated as strongly predictive features into a supervised classifier for applications such as short-term mortality prediction [34, 64], sequential diagnosis prediction [34], and first-time drug usage (e.g., insulin regime adoption by diabetic patients [61]). Given the interpretability of the topics, we can then use them to understand the target outcome based on their corresponding importance scores derived from the trained classifiers.…”

A graph-embedded topic model enables characterization of diverse pain phenotypes among UK biobank individuals

“…Several existing methods utilize topic models to find meaningful latent topics from electronic health record (EHR) data using structured administrative data such as the ICD codes ( Li et al., 2020 ; Song et al., 2021 ). In contrast, we demonstrated the utility of GETM on the less structured and more sparse self-reported questionnaire information from the UKB including 443 conditions and 802 medications.…”

“…Several topic methods were developed recently for effectively mining EHR data ( Li et al., 2020 ; Song et al., 2021 ). However, most existing topic models are unable to incorporate existing biomedical knowledge graphs, which manifest in several forms such as disease taxonomy and drug classification systems.…”