7月 5, 2018/Lauren Schivley/未分类, Decision Support

Artificial intelligence (AI) and machine learning are all over the news today and a priority for most data companies. Aktana’s product marketing director, Lauren Schivley, talks with Chief Science Officer Marc Cohen about the uses of machine learning, why it’s so valuable for decision support systems, and how it applies to commercial life sciences.

Let’s start with the basics — what is machine learning and how does it differ from traditional statistics?

That’s a big question, but put simply, machine learning is an evolved version of traditional statistics. Both ask the same question — how do we learn from data? — but have different ways of approaching it. For one, traditional statistics was derived from foundations in probability theory and mathematics, evolving from a time when computers didn’t even exist. Today, we can leverage the power of computers to expand these statistical theories through machine learning, which enables us to approach the question differently.

Machine learning also differs from traditional statistical methods in that it’s less interested in the drivers of a result and more interested in predicting what is likely to happen. Take the case of self-driving cars. No one worries about how the machine detects people in front of a car. As long as the machine keeps us safe by successfully predicting and navigating the world, we’re happy.

How is machine learning used in decision support systems?

Machine learning algorithms sift through tons of data in real-time to identify behavior patterns. The algorithms generalize these patterns across individuals and then predict which action to take on a person to elicit some desirable behavior. All of this processing happens behind the scenes as the machine collects data, analyzes it, and predicts the impacts of actions. The resulting insights are presented as recommendations to support decision-making.

In the commercial life sciences, machine learning has enabled sales and marketing teams to move away from segment-based marketing toward 1-1 marketing.

Let’s take a typical marketing task for reps — creating a messaging strategy. Previously, reps only sent messages based on HCP (healthcare professional) segment data. Today, machine learning platforms provide informed messaging recommendations for each individual HCP. The machine learning algorithm analyzes HCP data based on message history, previous reactions, demographic data, and any existing segment membership to identify patterns and recommend an action, such as which message to send, when to deliver it, and what channel to use.

Machine learning also helps sales and marketing reps do their jobs more efficiently. For instance, machine learning algorithms can improve rep efficiency by identifying reps’ travel patterns for certain territories and suggesting who and when to visit based on HCP availability.

That makes a lot of sense — and how does that fit with the brand strategy?

Great question. Moving toward a 1-1 marketing plan that completely relies on machine learning is a big step, especially for organizations that are used to traditional segment-based strategies. Aktana supports all approaches — the decision support engine can be based on the segment-based strategy, on the machine learning algorithm, or on some hybrid with elements of both.

The hybrid approach blends strategy-based inputs and machine-learning-based inputs to provide real-time optimization and automation of some aspects of messaging and sales execution, while adhering to constraints that are expressed in the brand strategy.

For example, sometimes teams want to deliver one or two email messages from a larger potential set of emails. In such a case, machine learning could analyze HCP past reactions to the message headers and from there identify the best email of the current set to send. This would be a hybrid approach — machine learning creates a recommendation within the context of the brand strategy.

Can you share more details about how “next best message” works?

Aktana’s machine learning technology can predict the most effective message sequence for a particular HCP in real-time. Here’s how it works. For every email message sent to a set of HCPs, the system analyzes the sequence of all other messages that were sent to each HCP previously, as well as their topics. This information is then used to generate a model that predicts the likelihood that an individual HCP will open a new message. The prediction is made for each potential message. Lastly, the rep receives a suggestion as to which email message will most likely be opened and can take action from there.

We recently published related research in The Journal of the Pharmaceutical Management Science Association, discussing how the Aktana platform uses machine learning to optimize message sequence. It’s a good read for anyone interested in getting deeper into the technical details.

Sounds like there are a lot of applications for machine learning. Which decisions are best left to the machine and how is this evolving over time?

Behavior patterns like the chance that someone is going to open an email or the best time to follow-up after an in-person visit are best identified and predicted by machine learning algorithms. With behavioral-driven decisions delegated to the machine, commercial leaders don’t have to focus on tactics but can turn their energy to strategy development and higher level direction setting.

We’ve already seen some evolution here with how clients use the Aktana platform to optimize message sequencing and sends. As I mentioned earlier, the machine can already predict the best next message to send given a set of existing messages. Our customers currently use this to aid the decision-making process, but as machine learning gets even smarter, they will be able to transition these messaging decisions to the machine entirely.

One final question — what are the biggest opportunities and challenges in machine learning?

There is a universe of opportunities, and Aktana has only just begun explorations. The challenge is to carefully assess each opportunity based on a clear business objective and a way to measure success. Sometimes that’s straightforward (like with “next-best-message”), but sometimes it isn’t.

Take, for example, the business objective of increasing script writing by identifying the best rep visit frequency. Visit frequency is easy to measure, but an increase in script writing may not be because of factors like lag times between treatments and results or the natural variability in scripts. That doesn’t mean it’s impossible to use machine learning to increase script writing, it just means that a lot of care needs to be taken in building the application.

A few other opportunities include using the machine to identify which reps are good at adhering to strategy (and which aren’t) and analyzing how the wording of email headers, suggestions, and insights changes adherence. Our text analyses have shown that word choices as features are predictive, so we hope to leverage this to make headers, suggestions, and insights more effective in driving behavior.

As you can tell, there’s a lot going on in machine learning. One thing’s for sure — it’s an exciting time to be working in this field.

ABOUT MARC COHEN

As chief science officer (CSO), Marc oversees analytic innovation at Aktana. He is an experienced technology leader who has spent much of his career improving business decisions using statistics, data mining, and machine learning. He spend half a decade as CSO at Archimedes, a healthcare modeling organization. He is an expert in modeling and applying algorithms to business problems, and has numerous publications and patents.