15 January 2020, 13:10
In this blogpost, we’ll explain how AI can help solve typical challenges in the commercial model of life sciences companies.
In an Accenture survey (1), more than 90% of life sciences executives recognized artificial intelligence as important in driving innovation and achieving outcomes such as hyper-personalized experiences, new sources of growth, and new levels of efficiency. As their market pivots to the patient with more individualized treatments and value-based models, life sciences companies must also shift their commercial models.
Traditionally, the lion’s share of pharma promotional strategy and investment has been focused on the interactions between the HCP (HealthCare Practitioner) and sales representative. Other promotional channels are meetings and events, service team calls, inside sales, digital, educational activities, etc. For sales organisations, it’s hard to measure the impact of each of these channels and how they influence each other. Also, defining the most effective channel mix for a specific HCP is not easy.
Through segmentation and targeting, commercial teams aim to tailor their efforts to groups of similar HCPs but this is often based on limited information, leading to inaccurate and too broadly defined HCP segments. No surprise that 1 out of 2 life sciences commercial leads said they don’t have a good understanding of what their customers need and want.
The AI revolution on the commercial side of the pharma business has been slower on the uptake than on the R&D side, but we see great opportunities to improve the commercial model through AI in many ways, of which 2 of them we’ll detail in this blogpost.
Segmentation & Targeting
Commercial organisations in Life Sciences often use HCP information like the number of patients treated for a specific disease, or the % of adoption to its product as a way to segment HCPs. A classical segmentation could be Gold-Silver-Bronze, with Gold referring to HCPs that treat more than 20 patients with a specific disease per week. Silver means 10 to 20 patients with that disease per week, and bronze means 0 to 10 patients. As a consequence, sales representatives are incentivized to visit Gold HCPs more frequently than Silver HCPs which they visit more frequently than bronze HCPs. This example shows that only a limited amount of information is used to segment the HCP population, and that a one-size fits all approach (at least per segment) is used. With AI, we can radically change this approach, and create a segment-of-one for each HCP. Curious to know how we do that?
Well, in order to address this we use what is called an embedding space. For the non-techies reading this post, please bear with me for just a few seconds. An embedding is a relatively low-dimensional space into which you can translate high-dimensional vectors. Embeddings make it easier to do machine learning on large inputs like sparse vectors representing words. Okay, so far for the definition, but how can this help me improve my segmentation and targeting you might ask. In essence this technique allows the use of all available data about HCPs, in any format (full text, database, time series, image, spoken text, etc.). Think of conversations, interactions with your HCP portal, click behaviour on your website, specialty, potential, adoption, interviews, medical facility, age, hobbies, geography, etc. For each of these ‘dimensions’ the embedding space maps the values (e.g. age number) on an axes to distinguish between HCPs. The illustration below gives 2 examples of what this looks like for a combination of 2 dimensions (e.g. gender and royalty in the left example).
The illustrations show this in a 3 dimensional space, as this is the maximum number of dimensions that can be visually illustrated, but in fact the number of dimensions used in the embedding space can grow to infinity. But let’s stick with the 3D visualisation for simplicity. So by doing this exercise with all data at hand, every HCP will be given specific coordinates in the embedding space. The closer that two HCPs are located in this embedding space, the more similar they are.
So for example, a general practitioner living in location ABC might appear to be very similar to a pneumologist in location XYZ, because they attended the same university together, are of young age so prefer digital channels, practice the same hobbies and both frequently attend conferences. These two HCPs should be targeted in the same way based on the information coming from the embedding space, whereas traditional methods would target them differently based on limited information (e.g. only potential).
We at ML6 applied and benchmarked this technique of hyper personalisation at a multinational company and outperformed the other techniques by 150%.
Analysis of the responsiveness of sales to promotional activities can be done through the smart use of data. Measuring brand sensitivity to promotion prior to investment decisions or considering implementation of new channels is a crucial step to maximize return on investment. Typical questions that arise are: “Which channels contribute significantly to the brand sales? What are my incremental sales per extra unit of investment? What is the optimal point of investment? What are the major sales drivers? How do sales drivers vary across regions or across promotional channels? What is the level of carry-over for a brand (base)? What is the optimal activity mix? etc.”
To answer these questions we make use of the unobserved components (UCM) time series model. This model was first introduced to the econometrics and statistics fields by A.C Harvey (1989). UCM can be considered to be a multiple regression model with time varying coefficients. It is based on the principles that it is useful to view time series as being decomposable into a trend, seasonal and cycle component.
Advanced modeling techniques (State Space modeling) are used to isolate, quantify and optimize the short-term impact of promotional activities on Sales.
Our model takes into consideration the carry-over effect (e.g. my weight this year is impacted by my weight at the beginning of the year (starting point) and how it evolved the years before), seasonal trends (e.g. ice cream sells better in summer than winter) and other known parameters (short and long term). It also accounts for the fact that part of the generated sales can be attributed to patients initiating treatment at a hospital continue using the prescribed drug after discharge from hospital, i.e. the hospital spill-over effect. This is all represented in the ‘Base’ (grey area in the model chart). This technique allows to account for the parameters that have influenced sales which we are not aware of or we do not have an accurate way to track them, e.g. competitor incentives. Therefore, sales impact is not misattributed to other channels such as e.g. traditional calls.
Next to that, the model also accounts for what is called the memory (ad-stock effect). This refers to the impact that marketing activities have over time on sales or brand health: It captures how response to advertising builds and decays in consumer markets. This concept agrees with common sense that the awareness level of a new exposure will be higher if there have been exposures in the fairly recent past and lower if there have not been. Finally, the delay in impact of an activity and diminishing returns over time are accounted for as well.
The model shows the impact of each promotional channel on sales, but can also be used to analyse responsiveness to additional investment in a promotional channel, which we will detail in a following blogpost.
The fact that science and technology are converging to enable more personalized, precise treatments for patients should also trigger sales and marketing professionals to apply similar techniques for more precise targeting and more effective commercial efforts. The current state of modelling techniques and hyper personalisation allows to radically improve sales and marketing operations for Life Sciences companies.
More information on how to apply AI in the life sciences industry can be found on https://ml6.eu/lifescience/
1 Accenture Industry X.0 Survey, Life sciences respondents, 2017 https://developers.google.com/machine-learning/crash-course/embeddings/video-lecture