It's beyond doubt that data is increasingly important in healthcare, but there is also a strong sense that doctors themselves are not very keen on it. I wrote last year, for instance, about a study exploring how doctors felt when patients brought their own data into consultations.
“We’ve heard doctors say more and more that people bring this data into the clinic and they’re just overwhelmed by it. When you’re managing chronic disease or symptoms, day-to-day lifestyle tracking data can be useful, but doctors don’t have a way to use these data efficiently and effectively,” the authors say.
This perhaps goes some way to explaining a finding from a second paper, published earlier this year, suggesting that doctors themselves are often the main barriers against the digitization of patient records.
A pair of papers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) aim to help doctors make better use of the digital information they're presented with.
The first paper documents a machine-learning based approach called ICU Intervene, which uses data from ICU to provide doctors and nurses with the best treatment given a range of symptoms. In addition to providing suggestions, it also explains its reasoning to provide a level of accountability that is so important in healthcare.
"The system could potentially be an aid for doctors in the ICU, which is a high-stress, high-demand environment," the authors say. "The goal is to leverage data from medical records to improve health care and predict actionable interventions."
The aim of the project was to provide actionable insights that can make a profound difference to the health outcomes of the patients on the ICU ward.
"Much of the previous work in clinical decision-making has focused on outcomes such as mortality (likelihood of death), while this work predicts actionable treatments," the authors explain. "In addition, the system is able to use a single model to predict many outcomes."
The system provides hourly predictions for five different interventions that aim to cover a variety of critical care needs, be that breathing assistance of the improvement of cardiovascular function.
Each hour the system pulls out data that allow for vital signs to be represented alongside the patient notes and other data points. The data is presented to staff in a way that clearly shows how far from the 'ideal' (average) the patient is.
Perhaps more importantly, the system is also capable of predicting the future. For instance, the system was able to accurately tell whether a patient would need a ventilator six hours into the future, which is a marked improvement on the 30 minute predictions available on many ICU wards.
Through testing, the system outperformed existing systems at predicting the correct intervention, and was particularly strong at predicting the need for vasopressors, which are used to tighten blood vessels, and therefore raise blood pressure. The team hope to continue developing the system so that it can support individualized care and advanced reasoning for the treatments it recommends.
Data transfer
The second paper documents a system known as "EHR Model Transfer", which is designed to facilitate the use of predictive models based upon an electronic health record system, even if it's trained on data from a completely different system.
This is a crucial piece of work as most machine learning models require data to be stored in a consistent manner. The carousel many hospitals have with constantly changing EHR systems significantly hinders attempts to utilize machine learning.
More successful data transfer is a big thing therefore. EHR Model Transfer works using natural language processing to identify key clinical concepts that may be encoded different across different systems. These can then be mapped to a number of clinical concepts.
The system is particularly useful when data needs to transfer between different systems, such as when patients move hospitals, but doctors still need to derive insights from that data.
"Machine-learning models in health care often suffer from low external validity, and poor portability across sites," the team say. "The authors devise a nifty strategy for using prior knowledge in medical ontologies to derive a shared representation across two sites that allows models trained at one site to perform well at another site. I am excited to see such creative use of codified medical knowledge in improving portability of predictive models."
The system was tested according to its ability to predict both the mortality risk of the patient, and their need of a prolonged stay. The system was trained on one EHR platform and then tested on a different one. It managed to outperform existing systems that only had to work on a single EHR system.
Both systems were trained using data from the critical care database MIMIC, which contains data from around 40,000 critical care patients.