Project Updates

Published work related to the project

Chaturvedi, J., Stewart, R., Ashworth, M., & Roberts, A. (2023). Distributions of recorded pain in mental health records: A natural language processing based study. BMJOpen https://bmjopen.bmj.com/content/14/4/e079923

Chaturvedi, J., Wang, T., Velupillai, S., Stewart, R., & Roberts, A. (2023). Development of a Knowledge Graph Embeddings Model for Pain. AMIA Annu Symp Proc. 2024 Jan 11 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10785867/

Chaturvedi, J., Chance, N., Mirza, L., Vernugopan, V., Velupillai, S., Stewart, R., & Roberts, A. (2023). Development of a Corpus Annotated with Mentions of Pain in Mental Health Records: A Natural Language Processing Approach. JMIR Formative Research. https://doi.org/10.2196/45849

Chaturvedi, J., Velupillai, S., Stewart, R., and Roberts, A. (2023). Identifying Mentions of Pain in Mental Health Records Text: A Natural Language Processing Approach. Studies in Health Technology and Informatics, 310, 695–699. https://doi.org/10.3233/SHTI231054 Awarded Best Student Paper at MEDINFO 2023.

Chaturvedi, J., Mascio, A., Velupillai, S. U., & Roberts, A. (2021). Development of a lexicon for pain. Frontiers in Digital Health, 3, 778305. https://doi.org/10.3389/fdgth.2021.778305

Epidemiological Studies

The plan is to use the data extracted from clinical text to answer various research questions that are of interest to the pain and mental health community:

What is the distribution of recorded pain in different diagnosis groups within mental health care?
Is there a gender difference in recorded pain experiences within mental health electronic health records (EHR)?
What is the distribution of common anatomical locations affected by pain within different diagnosis groups in mental health care?
Does recognition of pain in primary care influence (the use of) hospitalisations?

Exploring Pain

Conceptual Diagram of Pain

This conceptual diagram was built, using Grafo, after exploring 4 textual sources for how pain was mentioned - 2 electronic health records, 2 social media platforms

Google Trends for Pain

When a Google trends search was done for the word "pain" as a medical condition, and compared to other similar words like "fever" and "cough", the word "pain" showed most exponential increase in search over time. This highlights the community's interest in pain and justifies further researching topics around pain.

Annotating/Marking mentions of pain within clinical text

Annotation Schema/Plan

This is the plan that is being followed when annotating/marking information about pain within the clinical notes.

Examples of Pain Annotations

Here are some examples of how mentions of pain within the clinical text will be annotated/marked for further use in training a machine learning model

Summary of Annotations

5,644 annotations have been marked and this is a summary of their distributions across the different attributes.

Relevant annotations indicate mentions of physical pain afflicting the patient (such as "she suffers from headaches", "he has chronic back pain").

Negated annotations are when the mention indicates absence of pain (such as "no pain recorded", "patient is not in pain").

Not Relevant annotations indicate mentions not referencing the patient (such as "his mother was in a lot of pain"), or metaphorical mentions (such as "he was being such a pain") or hypothetical mentions (such as "she fears the pain it might cause").

Almost half of the relevant pain mentions had anatomical location or body part associated with the mention, such as "back pain", "tooth pain", "head ache".

Top 5 anatomical locations mentioned with pain

Only 11% of the pain mentions included the character of pain, such as "chronic pain", "constant pain", "burning pain".

Top 5 pain characters

Only 10% of the mentions included something about pain management measures, 7% of which were medications.

Top 3 Pain Management Measures

Study setup to answer research questions mentioned at the top of this page

A cohort of patient records were extracted from the CRIS database (more details about CRIS can be found in the "Resources" tab). This included patients who were 18+ years old and under an active referral at the South London and Maudsley NHS Foundation Trust between July 1, 2018 and July 1, 2019.

A machine learning based application for automatic identification of pain from text (this was developed using the annotations described in the previous sections) was run on the documents that belonged to this cohort of patients. This application helped identify which patients amongst the cohort had relevant mentions of physical pain in their clinical notes.

Q1. What is the distribution of recorded pain in different diagnosis groups within mental health care?

In the graph below:

SMI stands for Severe Mental Illnesses, and includes people with a diagnosis of schizophrenia, bipolar or depressive disorders.

0 indicates class 0 i.e. not relevant - no mention of pain, or negated mention of pain (absence of pain mentioned), or hypothetical/metaphorical mention

1 indicates class 1 i.e. relevant - mention of physical pain affecting the patient

Both groups show higher frequencies of recorded pain.

Q2. Is there a gender difference in recorded pain experiences within mental health electronic health records (EHR)?