Use of a Medical Language Processing-based approach

Use of a Medical Language Processing (MLP)-based approach to extract depression symptom severity and suicide ideation from clinical notes to support depression research

Background

Depression is a major public health concern, and new research is needed to improve patient outcomes.

Electronic health records (EHR) contain extensive data that could be used for depression research, but some critical data, such as depression symptom severity as measured by the Patient Health Questionnaire-9 (PHQ-9) and Suicide Ideation (SI) are recorded in clinical notes rather than in structured fields.

Medical language processing (MLP)* can extract relevant information from clinical notes and fill in the gaps on depression symptom severity and suicide ideation in real-world data sources to support research and clinical care.

Objective

This study aimed to determine the feasibility of extracting PHQ-9 scores and Suicide Ideation from clinical notes.

Medical language processing can extract relevant information from clinical notes and fill in the gaps on depression symptom severity and suicide ideation.

*Referred to as NLP in the original poster

Methods

Data Source

Data were drawn from the OM1 Real-World Data Cloud, a continuously updated dataset of US patients derived from deterministically linked, de-identified, individual-level health care claims, EHR, and other data covering 2013 to present day.

Two Study Cohorts

The PHQ-9 cohort was restricted to patients with at least one clinical note that mentioned the PHQ-9
The SI cohort was restricted to patients with at least one clinical note that mentioned suicide

An MLP-Based Approach

Extracted PHQ-9 scores, associated dates, and SI using MLP
Identified collections of linguistic patterns used to record PHQ-9 scores, associated dates, and SI
Subject matter experts validated the patterns for each task to ensure reliability of the approach

For SI, notes were categorized as one of the following:

Negations of SI
Affirmations
Neither negations nor affirmations
Unknown

Results

1.1 million clinical notes were identified for inclusion in the PHQ-9 extraction cohort. Phrases validated by subject matter experts were used to extract 735,267 PHQ-9 scores.

The distribution of extracted PHQ-9 scores is shown in Figure 1.

3.7 million clinical notes were identified for inclusion in the SI cohort. Phrases validated by subject matter experts were used to classify 2,088,144 clinical notes as having negation of SI and 25,255 clinical notes as having affirmation of SI (Figure 2).

Of the 2,113,399 notes with either negation or affirmation of SI, 1.2% were affirmations and 98.8% were negations.

The remaining notes in the SI cohort were classified as neither negations nor affirmations or unknown.

Fig 1: Distribution of Extracted PHQ-9 Scores

Fig 2: SI Note Categorization

Conclusions

PHQ-9 scores and SI can be extracted from clinical notes using a MPL-based approach.

The approach used in this study is a reproducible model for efficiently extracting concepts related to depression from unstructured clinical notes.

Future research should examine the performance of this approach in other data sources.

View the full poster here for more details and a list of references.