Nearly 350,000 patients spend more than three weeks in NHS hospitals in England each year. This week NHS England announced plans to try to reduce the number of long-staying patients by 25% ahead of the winter flu season. The plan appears to be focused on reducing admissions by increasing access to primary care, providing more support for care home staff and making greater use of emergency day cases. However, in the long-term, data science may be able to offer a more effective solution. Predictive models could help to more accurately forecast hospital admissions and discharge rates. This would allow health services to better tailor system capacity to match demand. If disease onset can also be predicted, then it may be possible to intervene in the primary care setting before the condition escalates to a point where hospitalisation is required. There are a number of interesting pilots that have been undertaken in hospitals around the globe with these aims.
In Parkland Health and Hospital system in Dallas Texas, predictive analytics models were developed to help improve the allocation of health service resources by directing them towards the highest risk-patients. The model helped to identify the patients at highest risk of readmission in near-real time by using data from their electronic medical records. These insights were then used to direct scarce health care resources to patients in most need, both in the hospital and after discharge. This therefore helped to reduce readmission rates. Originally used for patients with congestive heart failure (CHF), the model has been extended to a range of conditions including diabetes, pneumonia and acute myocardial infarction. It is estimated that the decline in readmissions among CHF patients saved the hospital an estimated $500,000.
The start-up Pieces Technologies has now been set up off the back of this initial research, and the company is using artificial intelligence and natural language processing to provide software to improve patient outcomes across healthcare and community settings. The introduction of this sort of software into NHS hospitals could help to better identify which patients are stable enough to discharge and which patients would be at highest-risk of re-admission if discharged early. However, consent from patients to utilise their electronic medical records data in this way may be required. Particularly, if a private company will have access to this data.
Nevertheless, there is a precedent in the NHS for the use of medical records data by a private company for modelling. In 2016, Google’s DeepMind set up a collaboration with the Royal Free NHS Trust in North London to develop the clinical app Streams. As part of this collaboration, DeepMind accessed healthcare data on 1.6 million patients to try to improve monitoring of patients with kidney disease. DeepMind is now focused on using AI systems to improve access to and increase the speed of care. Earlier diagnosis and treatment in this way, could help to reduce admissions by enabling conditions to be managed in a primary care setting before severity increases and hospitalisation is required.
In terms of projecting hospital admission rates, 4 hospitals in Paris are using machine learning forecasting in this area. 10 years’ worth of hospital admission records were analyse using time series analysis provided by Intel’s software engineers. The aim was to predict down to the hour the arrival of patients at each hospital. Even though some data such as medical cause of admission was restricted due to data protection laws, the model was still able to leverage basic admissions data to successfully forecast admission rates 15 days ahead. This allowed the number of staff to be increased in peak periods. So even with basic hospital admissions data, there is an opportunity to use data analytics to improve the forecasting of demand for hospital services without divulging more personalised data.
If data on local hospital admissions could be coupled with crowd-sourced data on infectious diseases then it may be possible to more proactively manage infection before symptoms escalate to a point requiring hospitalisation. In Australia, analysis tool EpiFX can predict winter flu outbreaks up to five weeks in advance. There are also a number of consumer-based tools that forecast flu including the CDC FluView app in the USA. Using data from social media sites such as Twitter has also shown promise in predicting the spread of flu in real-time.
If data on hospital admissions, patient electronic records and crowd-sourced sickness data from dedicated apps and social media are pooled together then projection of the onset of pandemics should become easier. This can ensure that the health care service are strengthened in times of need. Based on these forecasts, health services can be more proactive in responding to emerging disease outbreaks. This would go a long way in meeting the NHS’s goal of reducing hospital admissions. Safeguarding patient data and investment in technology and expertise in the field of data science are likely to be the key to succeeding in this ambition.