Predictive analytics creating far-reaching benefits of near ‘real time’ hospital mortality ratios
Monitoring deaths in our hospitals has become a standard part of the healthcare landscape, helping to assess the performance of our hospital trusts and the quality of their care. The most common method is the Hospital Standardised Mortality Ratio, a national indicator referred to as the HSMR.
It is considered to be a good way of understanding whether people are getting healthier, or our hospitals are getting safer; although the Covid-19 Pandemic has severely tested that theory in recent times.
The HSMR ratio is a ratio of observed in-hospital deaths against the number of expected in-hospital deaths. It is calculated from what is known as a logistic regression model based on a mix of case factors for each individual discharge. Each discharge gets assigned a ‘mortality probability’ between 0 and 1.
It is important to look at it this way because simply tracking the actual (observed) number of deaths is not sufficiently helpful. As any mathematician will tell you, if the rate at which people in the population died stayed the same, but the population grew, then there would be more deaths overall.
In general terms, the rationale for calculating death rates in our hospitals is that they can be used to measure hospital quality, and therefore help trusts to:
- reduce/improve mortality rates
- improve patient care/safety
- reduce avoidable variations in care and outcomes
There were approximately 8.1 million discharges from UK hospitals between 1 October 2020 to 30 September 2021, from which 239,000 deaths were recorded either while in hospital or within 30 days of discharge, according to NHS Digital. This includes deaths from other causes as well as deaths related to the reason for the hospital admission.
HSMRs are some of the most acutely monitored data available to health trust leaders and healthcare professionals all over UK.
Taking the time lag out of mortality monitoring
HSMR data is delivered three months in arrears. So, for example, figures issued in April 2022 relate to January 2022.
Work being done by The Health Informatics Service (THIS) for its host trust, the Calderdale and Huddersfield NHS Foundation Trust (CHFT), is replicating the HSMR in near ‘real time’ using predictive analytics – an emerging asset in healthcare. Intitial findings show an accuracy level of 85 per cent when compared to the national statistics, thus providing invaluable insight for CHFT’s medical directors, Mortality Surveillance Group, and medical review co-ordinators.
Oliver Hutchinson, THIS’ Performance Information Lead for Primary Care, who is leading the mortality ratio analysis, explains:
“This model will pinpoint the cases [of mortality] to review. I think medical directors will be really interested in this; it’s not something they’ll need to review daily but looking at it on a monthly basis will provide them with data they would otherwise be waiting three months to view when it gets released at national level. There are so many people involved in this process, the reach is going to be widespread. I even think clinicians who come and sit in on the Mortality Surveillance Group will be interested as well.”
Predicting a patient’s chance of survival
Anyone arriving at hospital is assessed on several criteria – these include a patient’s age, gender, exposure to deprivation, how they were admitted, comorbidity, whether they have a paliative care code and the actual diagnosis.
Based on the information gathered, predictive analytics can allocate an expected death calculation (EDC) ranging between 0 and 1 – a score approaching 1 indicating an increased likelihood of death, while a score towards 0 indicating almost certain survival.
Taking into account all known factors, the scores are worked out incrementally, so for instance, someone might be given an EDC of 0.376 – which can be expressed as a percentage, ie they have a 37.6 per cent chance of dying.
Oliver Hutchinson explains:
“We feed the factors into a model and that produces the expected death calculation. It’s been a very iterative process. We’ve had to work through different challenges of feeding in so much data into the predictive model. That is very complex, with something like a diagnosis there are up to 2,500-3,000 different diagnosis codes, for example. There’s been a lot of trial and error to get that information to run and produce some outputs without it taking an enormous amount of time.”
Pinpointing trends and variations
The benefits of near real time HSMR will enable the trust to pinpoint trends and variations before the national figures are published, as well as improve its ability to improve its clinical review process.
The EDC figures create a ‘league table’ of patients - those at the ‘top’ with scores at the high end of the ratio scale and those at the ‘bottom’ being patients with the better survival chances.
“The benefits are twofold. The first is to be three months ahead of the national publication, so that when the national publication does come out, we’ve had foresight which will dispel any undue concerns. I can share the data with the board, with the mortality surveillance group, which is where this information goes, primarily.
“The second is that it will help us pinpoint potential clinical areas to improve. For example, if you’ve got a patient that had a low likelihood of death and they actually died, we can look into that case and hold a clinical review to see if there is anything we can learn to prevent a death in similar circumstances in the future.”
Pre-covid the Trust, which runs Huddersfield Royal Infirmary and Calderdale Royal Hospital, reviewed 50 per cent of deaths, although that figure has fallen because of the pressures created by the pandemic.
However, Gawaine Carter, head of THIS’ Information Structure team, believes the data will allow the trust a clearer focus on which cases to review. He says:
“The trust has always reviewed a good proportion of deaths but there weren’t pointers to which ones they really should be focusing on to learn from. The difference now is that it helps to focus a limited resource. Prior to this, we didn’t know where to focus until way after the event.”
Another use for the system is to potentially shed light on errors in recording a patient’s medical history. For example, a clinical review into someone who died despite a low EDC could check to see if certain factors had been missing from their medical record, which would decrease their survival chances and therefore would have produced a different EDC.
Collating current and historic data
Predictive analytics can process and evaluate enormous amounts of historic and real-time information to create valuable forecasts, predictions and recommendations, on anything from individual patient care to wider public health trends.
Predictive analytics can assess tens of thousands of data points ranging from a patient’s condition on arrival at hospital, including whether they arrived in an ambulance, car or by foot, to their individual medical records and broader socio-economic or demographic information, such as their home postcode and ethnicity.
THIS’ Information Services (IS) team is collating data from the CHFT’s Cerner Milliennium EPR and other far-reaching clinical systems to gather information that then populates predictive R analytics software, an open-source software used for all kinds of data science, statistics, and visualisation projects. R programming language is powerful, versatile, and able to be integrated into business information platforms, to help users get the most out of business-critical data.
Gawaine Carter says:
“The amount of time and effort the Trust puts into focusing on and reviewing mortality rates is remarkable. It started more than a decade ago and hasn’t relented since, especially due to the obvious pressures of the pandemic. It has always been really high profile, and occupies a lot of clinical time, admin time and business intelligence time.”
The correlated data is stored in a data warehouse that is accessed via a front-end system developed by Gawaine and his team using Qlik Sense desktop platforms that can be interrogated in fine detail ranging from top line data down to individual patient details.
Hardware, software and bespoke solutions
Creating the real time HSMR is THIS’ latest predicitive analytics project. Two previous projects have focused on ICU admissions during the pandemic and, more recently, A&E admissions, to predict the percentage of patients arriving at A&E being admitted to hospital or sent home – currently running at an average accuracy level of about 85 per cent.
Having reliable, secure and intelligent hardware and software is a crucial requirement to enable healthcare organisations to take full advantage of predictive analytics.
It requires technical infrastructure such as a data warehouse and a portal to access workstreams supported by highly skilled and experienced analysts who can extract the valuable insights they can provide.
Specialists at THIS can advise on, and provide the full extent of of these key requirements to help integrate a bespoke predictive analytics system into NHS healthcare organisations. By partnering with THIS, predictive analytics can be easily co-ordinated to help reduce operational costs, improve patient outcomes, and increase the effectiveness of an organisation’s resourcing.
Contact us to find out how we can help you.