I am text block. Click edit button to change this text. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.
The team at D&D have recently started to build a Command Centre.
The vision is for this to be used by healthcare providers to monitor their performance, find where their flow bottlenecks are and get a general sense of how the hospital is performing. This has been an exciting journey for us!
Presently, we are working with several different NHS trusts to get this embedded into their operational hubs. However, it should be noted that this process is still in development, and the list of algorithms stated will / may be subject to change.
What makes this solution different?
Our solution is different to the traditional ‘what has happened’ command centre solution, as it embraces, and is amplified, by our ML and predictive models. These algorithms and models are built using our D&D data science framework and can be easily adapted to work with other solutions.
Embracing AI and ML, and adapting our products to suit, allows the command centre to tell the healthcare provider what may happen, as many products on the market do well at informing what has happened. This is where our product stands apart from other similar types of solutions on the market.
Furthermore, due to our team’s experience working for (and with) the NHS, the team at D&D know what the customer needs and wants from the product. In addition, our data science platform allows our data science team, at D&D, to customise the predictive models, therefore allowing the user to shape the product to their needs.
What makes this solution different?
Our solution is different to the traditional ‘what has happened’ command centre solution, as it embraces, and is amplified, by our ML and predictive models. These algorithms and models are built using our D&D data science framework and can be easily adapted to work with other solutions.
Embracing AI and ML, and adapting our products to suit, allows the command centre to tell the healthcare provider what may happen, as many products on the market do well at informing what has happened. This is where our product stands apart from other similar types of solutions on the market.
Furthermore, due to our team’s experience working for (and with) the NHS, the team at D&D know what the customer needs and wants from the product. In addition, our data science platform allows our data science team, at D&D, to customise the predictive models, therefore allowing the user to shape the product to their needs.
What are the components of the command centre?
The command centre is setup as a six-screen view, these screens show:
- Emergency department (front door) key measures. These range from ED attends, conversion to inpatients to LOS and time to be seen, for example;
- Inpatient view – this shows the current beds occupied by patients and various other measures related to inpatients;
- Discharge view – this shows how many patients have been discharged, how many will be discharged and why they have not been discharged i.e. awaiting a TTO (time to take out from Pharmacy), delayed transfer of care (DTOC), etc;
- Three screens are then dedicated to showing the patient level detail to supplement each one of the summary screens. This PTL (patient tracking list) gives the service everything they need to manage a patients progress through the system. Additional functionality is in development to trigger alerts based of tolerances derived through analysis and qualitative experience of the service.
The below image shows our CEO Orlando Agrippa launching our command centre prototype at D&D’s Executive Patient Flow Summit (https://www.draperanddash.events/):
This event helped shape what the command centre has metamorphosed into.
Please sign up for next years event if you are interested in learning about D&D’s developments in this space.
How is the command centre amplified with predictive analytics / machine learning?
The command centre, as previously mentioned, uses retrospective outputs from each pilot site (normally 3 – 5 years retrospectively – dependent on the data quality and trustworthiness). This data is then trained using a specific machine learning model to look for the features that will be important factors in making a future prediction from the model.
D&D’s command centre has a lot of different data sources, but from the team’s experience and engagement with our initial pilot sites, these were the indicators that needed a prospective view:
| Indicator | Method used | Type |
| Admissions Forecast | Hybrid forecasting method (auto.arima, ets, nnetar, stlm, tbats, snaive) | Time series prediction |
| A&E Forecast | Hybrid forecasting method (auto.arima, ets, nnetar, stlm, tbats, snaive) | Time series prediction |
| Cancellations | Random Forest | Machine Learning |
| ED Conversions to inpatients | Naïve Bayes | Machine Learning |
| Bed Occupancy | Hybrid forecasting method (auto.arima, ets, nnetar, stlm, tbats, snaive) | Time series prediction |
| Radiology Turn-Around Times | xgbDART | Machine Learning |
| Time to be seen | Random Forest | Machine Learning |
| ED visits per hour | Hybrid forecasting method (auto.arima, ets, nnetar, stlm, tbats, snaive) | Time series prediction |
| Stranded | K-means clustering, Naïve Bayes and Recursive Partitioning And Regression Trees (RPART) | Machine Learning |
| Readmissions | K-means clustering, Naïve Bayes and Neural Net | Machine Learning |
| LOS Prediction | Gradient Boosting Machine (GBM) | Machine Learning |
The command centre currently uses a mix of supervised ML methods and univariate time series methods. The following series of bullets explain why we have chosen these algorithms and explain how they work:
- D&D’s hybrid forecasting method is a collection (ensemble) of various time series models. Our algorithm chooses the model with the best accuracy and uses this to make future predictions. This works by using autoregressive (how previous and current observations correlate on a time series axis) patterns inherent in time series data. This method is then used to make predictions of inpatient admissions, A&E arrivals, bed occupancy, visits per hour and will be used with other measures. Please refer to https://www.draperanddash.com/machinelearning/2019/09/predictive-solutions-series-forecasting-module-and-toolset/ for more information on how the time series predictive solution is created and deployed.
- For cancellations and time to be seen we use Random Forests, these are a series of tree-based models that allow for visibility of where splits occur in the data. Random forests, or random decision forests, are an ensemble learning method for classification, regression and other tasks. These operate by constructing a multitude of decision trees (when training the model) and outputting the class (classification) or mean prediction (regression) i.e. predicted value. The reason machine learning is used here is to do with a more complex relationship with when a patient cancels (cancellations) and at what time the patient will be seen in ED (time to be seen).
- For ED conversions to inpatients our ML model uses a probabilistic classifier (namely Naïve Bayes) to generate predicted probabilities. These probabilities state the likelihood of a patient being converted from the emergency department to an inpatient setting;
- Radiology turnaround times – use the algorithm (xgDART) – this algorithm drops trees in order to solve the over-fitting. This drops less significant trees to correct errors in predictions and to adjust over fitting. This data set was prone to over fitting, thus this classifier was used to correct this issue.
- Stranded patients is informed by our existing algorithm (https://www.draperanddash.com/machinelearning/2019/09/predictive-solutions-series-stranded-and-super-stranded-ml-module/) and readmissions informed by https://www.draperanddash.com/machinelearning/2019/09/predictive-solutions-series-readmissions-ml-module/.
- LOS prediction – this uses a method called gradient boosting, with the aim to produce a prediction model in the form of an ensemble of weaker prediction models (typically decision trees). It builds the model in stages and generalises them by allowing the optimisation of a differentiable loss function. The choice of this model, related to LOS, is purely to do with increasing the accuracy of the predictions derived from the model. Therefore providing more sensible estimates of how long a patient will stay in hospital.
By utilising this combination of methods the command centre is informed by cutting edge algorithms, designed to solve most regression and classification problems.
We are on an exciting journey at the moment, but this is just the start of the journey to amplify the command centre with predictive analytics and machine learning. Through our engagement with our pilot sites, and over time, we expect the command centre predictions and components to evolve in line with the development of the solution.
Please enquire if this solution sounds like just what your healthcare provider needs. You can reach us at sales-support@draperanddash.com.