Identifying Optimal Physician Targets with AI Creates 34% Sales Increase

The client came to us wanting to utilize big data and AI to identify things such as ‘Which doctor has the most potential to write a script for a patient appropriate for our brand, today?’ and to help the rep understand ‘What should be the priority, based on the most recent data, to gain more scripts of our brand?’ and ultimately with channel data added, we should also be able to identify, ‘What messages and channels and sales and marketing actions will enhance that outcome’. In addition, with appropriate patient data we could also identify maximum share per physician based on their patient population.

The process we followed to achieve this had four main components of work.
1. Gather relevant and appropriate data
We need unique physician IDs as a base, then from that we layer on as much details about those physicians as we can from big data sources, such as claims data, Rx data, CRM data etc.
2. Sort, clean, transform and combine the data
This data wrangling component is always the most time consuming part of any project. Data in big data projects has to be combined. You cannot simply drag and drop as you can with data visualization programs (that are not analytics – see blog (The Difference Between Data Visualization and Data Analytics – and Why It Matters) and expect to get a sensible result. In data visualization programs without analytics you are not finding complex predictive relationships in the data as you would be with AI powered analytics. Once the data is cleaned, and coded (e.g. location data needs to be binary coded to be recognized in an algorithm while other types of data need to be normalized, features need to be selected (as leaving some features in would be detrimental to accurate results) and so on. Once the data set is ready it is typically divided into 3 parts; the training data set, the validation data set, and the test data set.
3. Create Artificial Intelligence algorithms that suits the needs of the data and objectives best
There are many techniques to choose from and even if using machine learning there are many different machine learning techniques to choose from. They could be decision trees, neural networks and many more. It is up to the data scientist to choose which techniques best suit the problem and data combination. Once they choose the technique there are many parameters to be created and tuned within algorithms to achieve strongest accuracy.
4. Train, Test and re-evaluate the model
Once the model is built it must be trained, and then re-evaluated for accuracy. Once a strong level of accuracy is attained, it gets tested on the test data set. That data set will have data unseen by the model previously so the accuracy can be assessed with new data. Getting the accuracy strong can be a time consuming process.