Technology Trends MIT's computer science and artificial intelligence laboratories use various types of medical data, including electronic health data records, to predict medical conditions. The two teams created the "ICU Intervention" and "EHR Model Migration" machine learning methods, respectively, to improve patient care conditions.

Doctors are often troubled by the need to look at various charts, test results, and other indicators. It is very difficult to make real-time treatment decisions while integrating and monitoring multiple patient data, especially when the data records between hospitals are inconsistent.

Researchers at the Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory (CSAIL) discussed in a new article how computers can be used to help doctors make better medical decisions.

Among them, a team created a machine learning method called ICU Intervene, which requires a large number of intensive care unit (ICU) data. The required data includes human vital characteristics and laboratory data. , note notes, demographic data to determine what treatment is needed for different symptoms. The system uses "deep learning" technology to make real-time predictions and learn from past ICU cases to suggest intensive care and explain the reasons for making these decisions.

Dr. Harini Suresh, the lead author of the ICU intervention article, said: "The system may help doctors who are on standby in the ICU because it is an environment with high pressure and high demand. Its goal is to use data from medical records. Improve medical conditions and make predictions about possible interventions."

Another team developed the "EHR Model Migration" method, which is capable of systematic training for processing data from different EHR systems and can be used to help apply prediction models to electronic health record (EHR) systems. Specifically, using this method of the research team, it is possible to implement a predictive model that trains mortality data and extended residence time on an EHR system, and migrates the model to another EHR system for prediction.

ICU interventions were jointly developed by Dr. Suresh, Nathan Hunt, postdoctoral Alistair Johnson, researcher Leo Anthony Celi, MIT professor Peter Szolovits, and doctoral student Marzyeh Ghassemi and were presented for the first time at the Boston Medical Machine Learning Conference this month.

The EHR model transfer was co-developed by CSAIL's doctoral students Jen Gong and Tristan Naumann, and Szolovits and Electrical Engineering Professor John Guttag. It was first presented at the ACM Knowledge Discovery and Data Mining Special Interest Group in Halifax, Canada.

Both models were trained using data from MIMIC, a key-care database that included de-identification data from approximately 40,000 intensive care patients and was developed by the MIT Institute of Computational Physiology.

Intensive care unit (ICU) intervention

Integrating ICU data is critical to automate the process of predicting patient health outcomes.

Suresh said: "Before this, many of the clinical decision-making work focused on the results of mortality, and the emergence of this work is to predict the possible treatment. In addition, the system can use a single model to predict a variety of results."

The ICU intervention focuses on forecasting the five key measures in hours, covering various critical care needs such as respiratory assistance, improving cardiovascular function, lowering blood pressure, and infusion therapy.

Every hour, the system extracts values ​​from the data representing vital signs as well as clinical notes and other data points. All data are expressed as a value, indicating how far the patient is from the mean (then assessing further treatment).

Importantly, ICU interventions can make predictions for the future. For example, the model can predict whether the patient needs a ventilator after 6 hours, and not just predict that the patient needs to use the ventilator after 30 minutes or 1 hour. The team also focuses on providing reasoning for model predictions and providing doctors with more insights.

Nigam Shah, associate professor of medicine at Stanford University, said: "Neural network-based deep neural prediction models are often criticized for their machine identity. However, these authors highly accurately predict the beginning and ending of medical intervention. And it can actually confirm the interpretability of the predictions it made."

The team found that the system is superior to previous ones in predicting interventions and is particularly good at predicting the need for vasopressin, a drug used to tighten blood vessels and increase blood pressure.

In the future, researchers will work hard to improve ICU interventions so that they can provide more personalized care to patients and provide more advanced predictions for decision making, such as why a patient may gradually reduce steroids, or why another patient may need to perform Endoscopy.

EHR model migration

Another important consideration in using ICU data is how it is stored and what can happen when the storage method changes. Existing machine learning models need to encode data in a consistent way, so hospitals often change their EHR system may cause major problems for data analysis and prediction.

This is where the EHR model migration comes in. This method is applicable to different versions of the EHR platform, using natural language processing to identify cross-system coded clinical information and then map it to common clinical information (such as "blood pressure" and "heart rate").

For example, a patient in an EHR platform may be converting a hospital and need to transfer its data to different types of platforms. The EHR model migration aims to ensure that the model can maintain its ability to predict patient conditions, such as long-term patient stay, or the possibility of death.

Shah said: "Machine learning models for medical treatment often have the disadvantages of low system external validity and poor portability between sites. These authors have designed a sophisticated strategy to use the acquired knowledge in the medical ontology. , thereby resulting in a mutually agreed expression between the two websites, which can help the model to perform well on another website after training on the website. It can be seen that creatively using encoded medical knowledge to enhance the prediction model Portability, I am very excited. "

Using the EHR model migration, the team tested its model's ability to predict two outcomes: mortality and long-term hospitalization needs. They trained the model on an EHR platform and then tested their predictions on different platforms. It was found that the EHR model migration was superior to the ordinary method, and compared with the EHR-specific events alone, the EHR prediction model was able to perform better data migration.

In the future, the EHR model migration team plans to evaluate the data and EHR systems of other hospitals and nursing organizations.