Browse and search developer information

Predictive Modelling Solution Framework

By Connecting for Health | 2012

Solution Framework

To understand how predictive modelling solutions can be developed and implemented a simple three stage framework is presented.

Figure 1 - Solution framework.

Figure 1 – Solution framework.

Stage 1 – Model Building or Selection

Some organisations may decide they want to build their own predictive model because; it will reflect the specific characteristics of their population, there is no available model for the event of interest and/or they want control over model maintenance.

To build a model they will need access to appropriate data sources and a method of analysing the data to produce a predictive model and test its accuracy.

As models are created from historical population data and the characteristics of populations change over time, it can therefore be assumed that a model once created will not stay valid indefinitely. Therefore any model will need to be recalculated or maintained over time. There is in fact a clear feedback loop between the effective use of the model and its validity. If the aim of the model is to predict and intervene to reduce the outcome event in the population, then the more successfully it is used the greater the impact on the population which in turn will reduce the accuracy of the model going forward.

Some organisations may not have the resources, skills or desire to build their own predictive model; instead they want to select an existing model.

Stage 2 – Prediction Tooling

Having built or selected a predictive model it must then be operationalised in the form of a predictive tool that can input the appropriate patient data, calculate the probability of the outcome and output it.

As for model building, some organisations may want to develop or commission their own bespoke tooling, while others may decide to use commercial or third party tools.

Stage 3 – Prediction Application

Once predictions have been made they need to be stored, managed and acted on by a clinician. This may result in interventions for a patient if appropriate.

Key to the success of any predictive modelling solution is user adoption, and in health care this is often driven by clinical leadership and adoption.

Conceptual Solution Architecture

Conceptually an end to end predictive modelling solution consists of:

  • Data sources that feed into a prediction tool
  • A prediction tool implementing a prediction model to calculate and output outcome predictions
  • Business applications such as risk stratification and case finding that process the outcome predictions

A prediction model is created by analysing data sources. Either an existing published prediction model, such as Combined Predictive Model (CPM), is used or a new bespoke prediction model is created.

Figure 2 - Conceptual architecture.

Figure 2 – Conceptual architecture.

Logical Solution Architecture

An example of the logical architecture of an end to end predictive modelling solution is:

  • Use of a Database (DB) or Data Warehouse (DW) platform to store and manage data feeds, predictions and business application processing of predictions.
  • Use of Extract Transform Load (ETL) DB/DW services to get and prepare source data ready for prediction processing
  • A bespoke prediction tool interfacing to the DB/DW.
  • Bespoke business applications interfacing to the DB/DW and also using a Business Intelligence platform to provide any required analytical functionality.
  • A portal front end to the prediction tool and business applications, which provides security and access control.
Figure 3 - Conceptual architecture.

Figure 3 – Conceptual architecture.

In many organisations physical services for DB/DW, ETL, BI and Portal will already exist.