AI is widely acknowledged as one of the most disruptive technology of our times that will transform every industry and function, and biopharma commercial is no exception. To avoid finding oneself behind others, biopharma commercial functions need to start acting fast on it. However, there is a lack of guidance on how and where to start.
While increasingly adopted in many industries and also some functions within biopharma, the adoption of AI/Machine Learning in biopharma commercial functions has been lagging. Multiple reasons include data regulations, lack of solid and multiple examples of success, other “high priority” business, low understanding, lack of executive support and so on.
Where AI/ML can help in commercial biopharma
As analysts, marketers, salespersons and other commercial functions in the biopharma industry, our overarching goal (and responsibility) is to make a difference in the lives of patients. Our work helps to bring life-saving treatments to patients. AI/ML can further help us to know where the patients are, who they are (classification) and when they need treatment (prediction), so treatments reach the right patients at the right time.
Broadly, AI/ML applications can be divided into classification and prediction problems. A few possible applications are:
Classification problems: Classifying disease or sub-disease where issues like misdiagnosis or lack of granular coding are prevalent, e.g. rare diseases, segmenting patient, physician and payers for better targeting
Prediction problems: Patient likelihood to switch therapy, adherence to treatment, physician likelihood to prescribe, predicting national / regional demand, and payer adoption
Having researched a lot of industry articles on this topic, I learned that, unfortunately, most articles don’t do a good job at guiding on how and where to start. This is one attempt in that direction so that we can realize the full potential of AI/ML faster.
As with any major change, the general process is to experiment, learn, and then deploy / expand. Start with a pilot. It’s very important to select the right problem, where chances of a positive impact are higher. A few rules of thumb to choose a pilot problem:
- Don’t start with the most complex problem.
- AI is data hungry - if not all, most approaches need a lot of data, and learning datasets (for certain types of approaches such as supervised learning). Ensure that you have sufficient data for the business problem that you are planning to solve, i.e. there are a good number and diverse set of variables or attributes (also called features in machine learning) which can possibly explain a response or outcome (also called target or label in machine learning) that is being predicted and thousands if not more observations.
- Most importantly, ensure that there is enough commercial value (or ROI (another story)) for the application, for knowing the unknown or predicting something. E.g. in several therapeutic areas, each patient on drug means an incremental revenue of >$50K, hence even if these approaches help in identifying a few patients every month, with good enough accuracy, it will add a lot of value. Real time data (e.g. weekly data) could be used to make this system run in real time for certain applications such as real time alerts to sales reps.
- Note that this is an experiment and the goal should be do this in 1–2 months, not a year. The longest part is getting all the data together, if that’s already there, this can be done very fast. Cost of doing this pilot should not be exorbitant.
Contrary to what is typically suggested, this experimentation doesn’t need to be a big initiative with an executive sponsor and a cultural shift. Instead of waiting for the organization to set an AI strategy, experiment, spend time and budget, managers can start themselves to learn and innovate. This could be part of their annual development plan or a stretch goal. The learnings can then be shared across the organization (who knows you might get a promotion as a result!!)
Here are a couple of examples from my own experience in learning and experimenting with new ideas. A few years back, when EMR data was relatively new, I explored it for a brand I was working with, and commissioned a study to find insights from unstructured EMR notes. The project helped me assess the data’s value as well as limitations. This was a time when approaches such as NLP were not as evolved as today. The project didn’t add as much value as expected, but it was fast and findings from it helped others in the organization who were looking to explore EMR data.
In another instance, I used spare end of year budget to do a patient segmentation using data beyond traditional claims, e.g. exhaustive demographic and lifestyle data for patients through a 3rd party source. We found that the additional patient dimensions were not very significant in predicting the drug choice for patients and adding them only provided small incremental benefit while increasing complexity, which was not desired. Again, great learning through a small and quick pilot, to assess the value of a new data source.
The AI/ML revolution is at our doorsteps and we need to act fast, for the business and for the patients. But we don’t need to wait for a major organizational change to start leveraging it. Following the approach highlighted in this article would help to take small steps which will then help to accelerate the learning curve.