Mining heterogeneous clinical notes by multi-modal latent topic model.

Area of Science:

• Computational linguistics
• Health informatics
• Bayesian inference

Background:

• Electronic health records contain valuable latent knowledge.
• Standard topic models fail to account for variations across different clinical note types (e.g., physician vs. nursing notes).
• This limitation hinders accurate extraction of patient comorbidities, symptoms, and treatment trajectories.

Purpose of the Study:

• To introduce a novel Multi-Note Topic Model (MNTM) for jointly inferring distinct topic distributions from heterogeneous clinical notes.
• To evaluate the performance of MNTM in improving topic interpretability and predictive accuracy compared to traditional single-note topic models.
• To identify patient health topics associated with adverse outcomes such as prolonged mechanical ventilation and mortality.

Main Methods:

• Applied latent topic modeling, specifically the proposed MNTM, to clinical notes from the MIMIC-III dataset.
• Trained MNTM to infer separate topic distributions for physician and nursing notes.
• Clinician-based manual assessments were used to evaluate topic interpretability.
• Correlated patient topic mixtures with clinical outcomes (prolonged mechanical ventilation, mortality).

Main Results:

• MNTM demonstrated significant improvements in topic interpretability compared to baseline single-note models.
• The MNTM model achieved significantly higher prediction accuracy for prolonged mechanical ventilation and mortality using limited patient data (first 48 hours).
• Identified specific diagnostic topics linked to poor patient outcomes.

Conclusions:

• MNTM effectively models distinct topic distributions across different clinical note types, enhancing knowledge extraction from electronic health records.
• This approach offers superior performance in both topic interpretability and clinical outcome prediction.
• MNTM holds promise for broader applications in analyzing diverse healthcare text data beyond clinical notes.