# Clinical Significance of Marital Status and Changes in Status Extracted from Unstructured Clinical Notes Using Ensembles of Off-the-Shelf Extraction Models

## TL;DR

This study shows how changes in marital status, extracted from medical records, are linked to health outcomes like depression and recovery rates.

## Contribution

The paper introduces ensembles of models to improve extraction of marital status from unstructured clinical notes and demonstrates its clinical relevance.

## Key findings

- Patients with marital status changes had higher depression rates compared to those with no changes.
- Marital status changes were also associated with higher recovery rates, indicating mixed health impacts.
- Ensemble models combining structured data and machine learning improved extraction accuracy of marital status from clinical notes.

## Abstract

Marital status and its dynamics significantly influence health outcomes. This study aims to demonstrate clinical utility of marital status and improve its extraction from electronic health records (EHR) by developing ensembles capable of using structured and unstructured EHR data.

Four ensembles based on different voting rules (unanimous vote, majority vote, precision-based vote, and vote based on random forest) were assembled using structured marital status data, off-the-shelf CNN, and hybrid models, capable of detecting marital status from clinical notes, followed by BERT-based fine-tuning to increase recall. Associations of marital status and its changes with depression, substance abuse, and other health outcomes derived from MIMIC-III were reported using frequency statistics.

Patients experiencing any marital status change (n=139, including getting married n=40), compared to patients with no change in status (n=2500) of similar age (M=57 years), had higher incidences of depression, but also higher recovery rates, suggesting dual health impacts of marital dynamics.