Measuring Brain Injury in Critically Ill Children: How AI Helps Enable Early Detection

Introduction

Each year, hundreds of thousands of children across America receive critical care. For these young patients, brain injury poses a serious risk that can lead to long-term health complications. Now, University of Pittsburgh Medical Center (UPMC) Children’s Hospital of Pittsburgh (CHP) has partnered with HAQM Web Services (AWS) to develop an innovative solution that can revolutionize the protection of these vulnerable patients. This breakthrough tool detects brain injury in critically ill children before any physical symptoms emerge. As a proof of concept (PoC), it demonstrates how integrating electronic health record (EHR) data with advanced machine learning can potentially transform patient care through early detection and intervention.

Medical Challenge

Critically ill children face a significant risk of brain injury due to factors such as toxins, reduced oxygen levels, increased intracranial pressure, and other less understood causes. Imagine the potential of identifying a brain injury before any visible symptoms appear. Traditionally, clinicians have had to rely on physical signs, such as a failure to wake or changes in pupil response. By the time these signs appear, the injury may already be well underway. We are using technology to tackle this challenge. Through the use of advanced analytics and machine learning, CHP and AWS have developed a solution called BRAIN AI. It uses an augmented intelligence algorithm designed to combine structured EHR data with novel brain injury biomarkers, paving the way for proactive interventions.

The Solution

CHP worked with AWS to build a data pipeline that ingests tabular EHR data and then uses the data to train AI models. The team at CHP deployed the trained AI model for exploratory analyses. The solution comprises two main work streams. The first workstream consists of data ingestion, data conversion, and machine learning model training.

Figure 1 – Data Conversion, Machine Learning

Here is an explanation of the steps depicted above:

1. EHR data in Comma Separated Values (CSV) format is sent from on-premises data warehouse to an HAQM Simple Storage (S3) bucket. These data include various patient clinical information, such as medications and vital signs. Fast Healthcare Interoperability Resources (FHIR) resource types used for this research consists of patient, observation, encounter, and medication administration.
2. AWS Batch, a job orchestration service, is initiated when files have been uploaded to the S3 bucket. The batch job consists of 2 steps. The first step is the execution of CHP provided Java code to convert the CSV medical data to FHIR R4 ndjson format, which is stored in another S3 bucket. The second AWS Batch step initiates the data import process by calling the AWS HealthLake API.
3. The imported data is accessible via FHIR APIs from AWS HealthLake, it can also be queried using HAQM Athena with the help of AWS Lake Formation and AWS Glue.
4. With the ability to use SQL to query the HealthLake data via HAQM Athena, the data science team performs data discovery, prepares the training data and trains AI models using HAQM SageMaker AI. SageMaker AutoML is used to automatically find the most optimal model and hyperparameters for this type of prediction. The selected model can then be deployed as a SageMaker endpoint for inference.

The second workstream represents the inference processing. In this example, inference can be completed in batch or in real time. The end goal of this solution is to hook directly into the EHR system for real-time inference.

Figure 2 – Inference Workstream, Batch and Optional Realtime

Here is an explanation of the steps depicted above:

Path – Batch Inference

1. EHR data is stored on S3 for batch inference.
2. HAQM SageMaker notebooks are used to process S3 data for inference.
3. HAQM SageMaker notebooks pass data for inference to HAQM SageMaker Inference Endpoint.
4. HAQM SageMaker inference endpoint, queries AWS HealthLake for additional features required for inference.

Path – Real Time Inference

A. To support real-time inferencing, an HAQM API Gateway and an AWS Lambda function is setup to handle inferencing requests. Incoming events, which can be FHIR requests, are processed by the Lambda function, which passes EHR patient and observation to the HAQM SageMaker Endpoint. The returned result is formatted as FHIR response back to the caller.
B. The HAQM SageMaker Endpoint, utilizes incoming event data and queries data from AWS HealthLake to compile features for inference.
C. AWS HealthLake contains patient and observational data used for inference.

Security

All services used in this solution are HIPAA eligible. The solution is fully capable of managing protected health information when deployed in a production environment. Data is securely transmitted from on-premises systems to AWS over a VPN tunnel, with encryption using Transport Layer Security applied as it is written to S3. Additionally, AWS Key Management Service encrypts data at rest. Comprehensive monitoring of all API calls and logging is provided by AWS CloudTrail and HAQM CloudWatch.

Next Steps

CHP plans to take this solution to production to refine their AI/ML models with production data. In the short term, the solution will work in batch mode based on daily data feeds. Ultimately, the solution will be integrated into patient care systems in real-time. This will help ensure brain injury can be detected more quickly, allowing professionals to treat children with greater accuracy.

The successful implementation of this PoC demonstrates how healthcare organizations can leverage AWS services to improve patient care through advanced analytics and machine learning. As CHP moves toward production deployment, this solution has the potential to significantly impact pediatric critical care outcomes.

Read on for more healthcare stories. To learn more about AWS for Healthcare & Life Sciences—curated AWS services and AWS Partner Network solutions used by thousands of healthcare and life sciences customers globally—visit the AWS for Healthcare & Life Sciences and AWS Healthcare Solutions webpages.