Extraction of LVEF from medical chart text

Output (actionable analytics)
Input(sample medical chart image)
Output (A json file with all extracted patient's LVEF values)

Inspiration and Introduction

I have worked in the healthcare analytics space for about 10 years. I have experience working for a large healthcare provider in the analytics and data architecture departments and also had the opportunity to develop NLP software for providers and payers. While I have a seen a tremendous amount of progress during this time, I still see many areas where advanced analytics could be utilized to improve patient outcomes.

One of the most underutilized data assets in healthcare is free text written in clinical notes. Developing systems to better harvest and operationalize this natural language data could result in earlier interventions, better health outcomes, and decreased healthcare costs. While highly data rich, this text is notoriously difficult to understand and process for NLP software. Amazon Comprehend Medical advanced entity recognition and relationship extraction could be utilized to process this text.

I use Amazon Comprehend Medical to detect the LVEF of heart failure patients and identify care improvement opportunities. Heart failure is the leading cause of hospitalizations and costs $216 billion per year in the USA (CDC). This project focuses on identifying an essential measure of heart failure patients, left ventricular ejection fraction (LVEF). LVEF measures how well the heart pumps blood and a low LVEF value means a patients’ heart failure is very advanced. In a healthy person, LVEF generally ranges from 55% to 75%.

One study found the LVEF is often buried in medical free text and which can result in patients missing out on life-saving care (Chaggar 2020). Another study (Kahn 2022) showed that a manual clinical review led to a 47% increase in the number of patients that were missing heart failure billing codes on their medical record. Additionally, 27% of missed heart failure patients were eligible for complex cardiac implantable electronic devices and another 45% required medication changes. Identifying and treating these patients would improve patient outcomes and increase revenue for the hospital system providing care.

Previous work to automate extraction of LVEF (Kim 2017) has some promising results, but further technological systems are needed to expand this commercially. This project will utilize AWS for image OCR (Textract) and Amazon Comprehend Medical for entity recognition and relationship extraction.

What it does

I built a system to extract the LVEF from medical charts and then provide analytics on the findings. There are two ways to use this product:

Input medical images
Input a csv with one column containing text from medical notes

Every healthcare provider should be able to easily access their medical records in one of the above ways.

The output is a json file containing the extracted LVEF for each patient/image and summary statistics on the results.

Example output summary

How we built it

We utilized a rule based extraction engine paired with Amazon Medical Comprehend and Textract. For any images being processed, we used Textract to OCR the image and saved the results as a text file. We then developed a pre-processing function to parse the text data prior to sending it to Amazon Medical Comprehend in order to extract the LVEF. We then parsed the results from Amazon Medical Comprehend through a custom post-processing function to get clean data. In order to evaluate the effectiveness of this solution, we tested the extraction on this synthetic dataset after manually labeling the correct LVEF for each patient. We then fine-tuned our pre and post processing functions based on the results we output from the synthetic dataset.

Challenges we ran into

While Amazon Medical Comprehend does an excellent job at extracting information for text, medical charts do not contain clean text data. They consist of duplicate notes (copied over from previous encounters) and often have excess informational data in the chart. For example, if a medical note contained the text Left ventricular ejection fraction in the range of 53% to 70% is considered normal, LVEF might be extracted as 53% to 70%. A human annotator of this sentence would likely understand that this is referring to LVEF in general, not the specific patient. We were able to mediate this challenge through the use of a pre-processing function where these types of sentences where filtered out. This function utilizes custom developed regex lookup queries and was developed through the error analysis of the synthetic dataset.

Accomplishments that we're proud of

Utilizing our synthetic dataset of 5000 patient discharges, we were able to successfully extract the LVEF for 93 patients. Additionally, 23 of those patients had LVEF values below 40% which means they would qualify for the I50.20 ICD billing code and 7 of those patients has severely reduced LVEF and might quality for cardiac implantable electronic devices. This will help hospital increase revenue and doctors to identify and treat patients in desperate need of care.

If these above results are extrapolated to a larger health system seeing around 250,000 patients a year assuming similar ratios to (Kahn 2022), we can estimate this solution's impact to health system and its patients:

~40 patients receiving life-saving surgery to get cardiac devices ₁
Additional ~$2.1 million in hospital revenue ₂

_{1. ((7/5000) X 250,000 X 27% X 47%)}

_{2. ((23/5000) X 250,000 X 47% X 240 USD) [billing codes] + ($50,000 X 40) cardiac surgeries}

What we learned

Amazon Medical Comprehend can be utilized to extraction complex clinical concepts from medical notes. However, its best use cases in healthcare will require significant pre and post processing of results to be clinically or commercially useful. We developed these functions based on synthetic dataset and tuned our models to optimal performance. Future work utilizing Amazon Medical Comprehend will need to be developed in a similar manner.

What's next for Extraction of LVEF from medical chart text

We recommend eventually licensing this technology out to large providers networks. The consolidation of the healthcare market in the past 10-20 years has led to the rise of fragmented databases and coding documentation. Unfortunately, patients have suffered from this by "slipping through the cracks" and not receiving they care they need. This technology could be used to rapidly iterate through provider's proprietary healthcare databases to extract structured LVEF data and obtain novel insight to their heart failure patient population. A refined list of patients with LVEF values would improve billing codes and identify patients that need referral to cardiology for advanced heart failure treatment, who might otherwise be forgotten.

A next logical step for this project would be to enter a partnership with a larger healthcare provider network. This would give us the opportunity to tune our code on a real dataset and allow the health system to receive novel insights to their heart failure cohort. The results could lead to an academic publication so we would target academic health systems for this partnership. I can utilize my connections to set up a partnership.

Advanced heart failure patients often result in complex surgical intervention which is generally how large health systems generate a large share of their revenue. This solution improves the quality of patient care and there are financial incentives for hospitals to implement it.

Citations

Matthew Kahn, Antony D Grayson, Parminder S Chaggar, Marie J Ng Kam Chuen, Alison Scott, Carol Hughes, Niall G Campbell, Primary care heart failure service identifies a missed cohort of heart failure patients with reduced ejection fraction, European Heart Journal, Volume 43, Issue 5, 1 February 2022, Pages 405–412, https://doi.org/10.1093/eurheartj/ehab629

P Chaggar, A D Grayson, N Connor, C Hughes, P1489 An audit of 215,000 patients on primary care registers using novel electronic searches to identify patients with heart failure requiring treatment optimisation and complex device therapies, EP Europace, Volume 22, Issue Supplement_1, June 2020, euaa162.332, https://doi.org/10.1093/europace/euaa162.332

Youngjun Kim, Jennifer H. Garvin, Mary K. Goldstein, Tammy S. Hwang, Andrew Redd, Dan Bolton, Paul A. Heidenreich, Stéphane M. Meystre, Extraction of left ventricular ejection fraction information from various types of clinical reports, Journal of Biomedical Informatics, Volume 67, 2017, Pages 42-48, ISSN 1532-0464, https://doi.org/10.1016/j.jbi.2017.01.017.