Yorkshire Ambulance Service worked with Hippo to understand how advanced analytics, geospatial modelling and an open-source digital twin could help test operational decisions.
Ambulance services make complex decisions under pressure. Demand, geography, workforce capacity, handover times, shift patterns and clinical prioritisation all influence how quickly a response can arrive. For Yorkshire Ambulance Service NHS Trust (YAS), covering dense urban areas and rural communities across the region makes that operational picture especially varied.
Serving over 5 million people
Employing 7,500 staff
And 1,000+ volunteers
YAS already had strong operational knowledge and rich data. The opportunity was to turn that data into explainable evidence that could support decisions about standby locations, triage scenarios, shift patterns and meal-break windows, while recognising the reality of frontline operations and the importance of crew welfare.
"Our engagement with Hippo has strengthened our analytical capability and established new opportunities to explore service improvements through data-driven insight. Hippo brought strong technical expertise and worked closely with our team throughout the engagement. The platform has strengthened our analytical capability by creating a transparent environment in which our data teams can safely test operational hypotheses, such as triage models and shift patterns, and explore potential impacts before changes are considered within frontline services. Equally important was Hippo's commitment to developing capability within our own teams, leaving us with both a valuable platform and the skills required to continue building on it."
Samantha Robinson, Chief Information Officer, Yorkshire Ambulance Service
Problem
Separating operational signal from anecdote
To understand the importance of this work, it is helpful to understand the national standards for ambulance response times in the UK. These separate categories have target response times attached to them and are based on the clinical urgency of the call-out:
- Category 1 (life-threatening): Mean response time of 7 minutes
- Category 2 (serious conditions): Mean response time of 18 minutes
- Category 3 (urgent): 90th percentile response time of 120 minutes
- Category 4 (less urgent): 90th percentile response of 180 minutes
YAS wanted to move beyond descriptive reporting. Existing data could show what had happened, but it was harder to understand what influences demand and response, and harder still to test how a potential change in operational policy might affect future performance.
Operational teams had experience-led hypotheses about the factors involved, including hospital handover, capacity, demand mix and crew availability. The challenge was to quantify those factors in a way that was explainable and reproducible for both technical and non-technical stakeholders.
This also meant working carefully with sensitive data. The project needed to bring together operational calls, resource records, capacity information, shift data, station locations and standby information without moving pseudonymised NHS operational data outside the agreed governed environment.
Using OpenStreetMap to identify isochrones of incidents with driving distances. Recreates manual process and using licenced software with open source.
Hexagonal grids enable more efficient summarisation of location data. This uses the H3 indexing method. This method enables faster processing, and makes it more suitable for real time decision making
Optimisation techniques to identify the best standby locations - these aligned with expected. It is challenging to assess if the optimisation is ‘correct’ as we can’t measure an effect on response times
Solution
Building a complex environment for risk-free testing
Hippo set up a governed AWS-hosted analytics environment, keeping pseudonymised YAS data inside the agreed environment while legal, security and information-governance activity progressed. Raw datasets were loaded to S3, converted to Parquet, catalogued with AWS Glue and made available for repeatable analysis through SageMaker notebooks and query tools.
With that foundation in place, the team explored response-time distributions, demand by area and time, conveyance patterns, station and standby coverage, and unused standby locations.
Operations research techniques helped frame standby planning as a facility-location and maximum-coverage problem, giving YAS a clearer view of coverage trade-offs.
The modelling was designed to be explainable. Feature-importance and SHAP analysis helped show which factors were associated with Category 2 performance, including missed meal breaks, call-category ratios and capacity shortfall. Where the data did not support a reliable model against the national Category 2 average target, the team reframed the analysis rather than over-claiming from an unsuitable model.
Building a digital twin
The aim was not to create a perfect replica of live operations. It was to create a practical, transparent environment where YAS could compare scenarios safely, understand trade-offs and decide where further operational testing or consultation would be most valuable.
The core output was an adapted open-source discrete-event simulator based on ELASPY (Electric Ambulance Simulator Python). With that framework, the team worked with the subject experts at YAS to create an event flow which aligned to a process their crews would take over the course of a shift.
A mix of previous work and understanding, combined with additional analysis, created the features of the model, including: shift patterns, planned meal-break windows, call categories, priority queues, conveyance to hospital, handover time, and en-route reassignment to closer ambulances when they became available.
Outcome
Evidence-based improvements in response times
The collaboration gave YAS notebooks, simulation code, scenario outputs and a senior-stakeholder playback. More importantly, it gave the team a reusable evidence base for exploring operational questions before committing time and effort to live trials.
In the simulator, a scenario in which 10% of demand that would otherwise enter Category 2 was clinically triaged into Category 3 was associated with a 15.9% reduction in median Category 2 response time and an 18.6% reduction at the 75th percentile. Separating meal breaks into two windows reduced simulated Category 2 response time by 13.9% to 19.4% compared with a single break window. Reducing missed meal breaks by 50% showed a modelled potential improvement of up to 3 minutes in South Yorkshire, 4 minutes in Humber and North Yorkshire, and 4.5 minutes in West Yorkshire.
Standby optimisation also showed how analytics could inform planning decisions. The modelling indicated that 80% coverage could be achieved with a median of 22 standby locations, with median maximal coverage of 96.5% at around 42 locations.
These figures are modelled scenario outputs rather than measured live-service improvements. Their value is in helping YAS understand where change may be worth exploring further, with a clearer view of the operational trade-offs involved.
Refined clinical triage
By triaging 10% of Category 2 demand into Category 3, the simulator indicated a significant drop in response times.
15.9%
Median response time 18.6% 75th percentile (reducing long waits)
18.6%
75th percentile (reducing long waits)
Smarter break windows
Moving from one meal-break window to two separate windows allows for better ambulance coverage. Modelled Improvement: 13.9% to 19.4% reduction in simulated Cat 2 response times.
