Hospitals in Paris are trialling Big Data and machine learning systems designed to forecast admission rates – leading to more efficient deployment of resources and better patient outcomes.
It’s just one more way in which cutting-edge data science is being applied to real-world problems in healthcare, along with creating personalized medicines, fighting cancer and streamlining pharmaceutical trials.
At four of the hospitals which make up the Assistance Publique-Hôpitaux de Paris (AP-HP), data from internal and external sources – including 10 years’ worth of hospital admissions records has been crunched to come up with day and hour-level predictions of the number of patients expected through the doors.
The core of the analytics work involves using time series analysis techniques – looking for ways in which patterns in the data can be used to predict the admission rates at different times. Machine learning is employed to determine which algorithms provide the best indicator of future trends, when they are fed data from the past.
The system is built on the open source Trusted Analytics Platform (TAP) – which was chosen for the task due to its capacity for ingesting and crunching large amounts of data, as well as its gearing towards open, collaborative development environments.
Kyle Ambert, a data scientist with Intel who contributes to the TAP project and worked with AP-HP on their implementation, told me “What was interesting with this work is that although there are many analytical solutions for these types of problems, none of them have been implemented in a distributed fashion.
“Because we’re interested in scalability, we wanted to make sure we could implement these well-understood algorithms in such a way that they work over distributed systems.”
The experimental project was carried out at four of AP-HP’s facilities, with plans to eventually roll it out to all 44 if it proves successful. Upscaling the project would increase the amount data which in turn would be likely to need the storage and processing capabilities of a distributed, cloud-based system, so a decision was taken to implement this model from the start.
This posed a challenge because, as Kyle told me, “With a lot of these problems there’s an off-the-shelf solution that’s extremely obvious and available. But there’s really no scalable implementation of time series analysis in the open source community.”
In order to complete the job, this meant Kyle and the team had to first build the tools they needed. The result was the first contribution to an open source framework of code designed to carry out the analysis over a scalable, distributed framework. This code is already being put to use in several other projects involving healthcare and finance.
Another obstacle were the French laws concerning privacy and patient data, which are relatively strict, meaning some data which would have been available if the system was running in the US, for example, was not available.
“There is other data we would have liked to have used – for example types of admission – was it heart attack, cancer, transfer from another hospital – but we couldn’t get the data into a format where it would have been useable and acceptable under the privacy laws,” Kyle says.
“But I was surprised by how well we could build the model without access to some of the data I would have liked. Just by leveraging patterns across multiple sites we were able to build a really good model.”
As well as the hospital’s internal data, several external datasets such as weather, public holidays and flu patterns were tapped.
The result is a web browser-based interface designed to be used by doctors, nurses and hospital administration staff – untrained in data science – to forecast visit and admission rates for the next 15 days. Extra staff can be drafted in when high numbers of visitors are expected, leading to reduced waiting times for patients and better quality of care.
The system has not yet gone into general use across the group but even if it remains an experiment, it has already achieved one of its objectives: bridging the “cultural gap” between hospital clinical and administrative staff, and data scientists and systems architects of the TAP deployment team.
Lessons learned will undoubtedly prove valuable for the group’s next Big Data project – building a data warehouse to store all of its clinical data in a form that can be interrogated by common techniques such as Python or R algorithms, while still complying with the sturdy EU data governance rules.
With the cost of providing healthcare increasing at more than the rate of GDP in every developed country, smart, intelligent systems like AP-HP’s are likely to play an important part in the future of healthcare. This is true in countries like France where healthcare is nationalized and funded through tax, just as it is in the US with its system based on private insurance contributions. By more accurately predicting the demand for services, waste can be cut and insuring can become more efficient. Ultimately, reducing the cost of healthcare is likely to lead to longer and happier lives for everyone.
