The Ultimate Guide to AI and Medical Insights

There is a lot of talk these days about artificial intelligence (AI), machine learning (ML) and natural language processing (NLP). What do these terms actually mean? What’s the difference? And more importantly what do they mean for medical/scientific insights? How will they impact Medical Affairs organizations?

In this ultimate guide for AI and medical insights, you will learn:

  1. The difference between AI, ML, and NLP

  2. What this means for analyzing medical insights 

  3. How AI can help Medical Affairs make better decisions

  4. The downsides of AI

What is Artificial Intelligence (AI)?

Artificial intelligence (AI) is a blanket term for computers making decisions like a human. defines AI as "the capacity of a computer to perform operations analogous to learning and decision making in humans…" This means ML and NLP are forms of AI:

Figure 1: Diagram of how artificial intelligence (AI) relates to machine learning (ML) and natural language processing (NLP). AI is a general term for applications that help computers “think” like humans. ML and NLP are subsets of AI.

Traditionally, AI meant giving computers a set of rules to follow in order to make recommendations or select outcomes. Think of this as using a decision tree for logic: "Hey computer, if this happens, then select this outcome."

Example of rule-based AI

Let’s say we want to make an app using AI to help people select their correct clothing size. The user enters their height and weight and the app makes a suggestion on which size they should buy.

With rule-based AI, we would determine a set of rules for the app to follow in order to make a recommendation (e.g., if over 5’6” and 180 pounds, buy a size Large. See Figure 2). The user would enter their height and weight and based on the rules programmed in, the app would spit out a recommendation:

Figure 2: Example of rule-based AI. Making predictions based on a decision tree. Based on height and weight, the model recommends a clothing size to the user.

Rule-based AI is great for many applications, but what if you don’t have a set of rules? Or it's too challenging to develop a set of rules? This is where ML comes in.

AI vs. Machine Learning (ML)

In contrast to traditional rule-based AI, ML lets the computer analyze a set of data, learn by identifying patterns and associating these patterns with a particular outcome, and then make a recommendation.

“At its core, machine learning is just a thing-labeler, taking your description of something and telling you what label it should get.” - Cassie Kozyrkov from Google

Instead of engineers building a set of rules for the computer to follow, you give the computer examples and say go figure out the rules yourself. This is a big deal for computer science because it means that humans no longer have to develop large complex sets of rules for computers to follow. Now, we can tackle problems we couldn’t address previously by giving computers examples of what we seek as outcomes. 

People often use the terms AI and ML interchangeably. Technically, ML is a subset of AI. 

ML example

In our clothing recommendation app example, instead of building a set of rules, we feed the computer a large dataset that contains people’s height and weight measurements and the size they bought. The computer analyzes and learns from the dataset in order to determine appropriate recommendations.

Figure 3: Example training dataset for ML. In the example of an app to recommend which clothing size to buy, an ML approach would utilize a large dataset with features or inputs (height and weight) with the known prediction targets (the size) to determine the best recommendations. Instead of humans creating rules for the decision tree, the ML models infers the best set of rules to make recommendations based on a large dataset (i.e., the ML models figures out the best rules). The more the data, the better the prediction model. An application for medical insights is to use ML to categorize the type of insight (such as competitive product, safety, etc). For this to happen, a large dataset of “labeled” or “pre-categorized” insights is required for the ML model to learn.

ML and Medical Insights

Ok, so ML can be used to label things. What does this mean for medical insights? MSL teams often capture several hundreds of insights per month and need to identify trends across insights. This can mean manually reading each insight and tagging it by hand. With the appropriate label, you can understand how frequently a topic is coming up and create graphs to analyze the trends over time. 

Figure 4: Example of a ML model automatically identifying top keywords in an insight and graphing over time. This avoids the time-intensive task of tagging insights by hand and helps Medical leaders quickly identify trending topics.

It’s a lot of work to do this by hand. And as you can imagine, it's highly susceptible to human bias. For example, different MSLs may tag things differently. Some MSLs may be very diligent in their tagging, while others may be...less so. Inconsistently-labeled insights can mean missing important trends or not getting the complete picture. Instead of doing this by hand, ML models can label insights for you automatically. 

For example, the automatic identification of a clinical trial name from insights can uncover which studies are discussed most in the field. Additional relevant keywords, such as Safety or Efficacy, can be identified to quickly understand HCP perceptions. Sentiment analysis (more about this below) can be applied to these insights to determine if a study is generally perceived to be positive or negative. With this information, Medical leads can determine the best next steps to drive the medical strategy forward. 

Benefits of Using ML for Medical Insights

  1. Saves time

  2. Applies consistent labeling to insights

  3. Reduces human bias

  4. Automates tasks

  5. Can be applied to other Medical Affairs data sources, such as MedInfo call records

What about natural language processing (NLP)?

NLP is about computers understanding and processing language like humans do. Think about what this means for a second...the complexity of human language is astronomical. There are so many nuances about language that we just know intuitively: accents, dialects, emotions, slang, how to make sense of spelling or grammatical errors, when different languages are combined ("Spanglish"), etc. How do we teach a computer all these nuances? The answer is it’s hard, but the field is making good progress. 

NLP models can be rule-based, use ML or be a combination of both methods. Learn more here.

Examples of NLP

A popular example of NLP is sentiment analysis. A computer looks at a block of text and determines the "mood" of the text. Is it positive, negative or neutral? A lot of sentiment analysis has been done on social media posts to understand how people "feel" about a particular topic such as politics or sports. Learn more about sentiment analysis here

Another example of NLP is keyword extraction. This method determines the main topic of the text and labels it accordingly. It's a useful approach to determine the relevant subject of a block of text. 

NLP and Medical Insights

Here are a few examples of how NLP can make MSL teams lives easier.

Capturing Insights via Voice

MSL teams can benefit from the voice recognition/translation aspect of NLP. Busy MSLs on-the-go can easily record insights while driving or in between meetings. This means that MSLs can easily enter insights while they're fresh and in less time.

ai and medical insights voice capture.jpg

Sentiment Analysis

Automatic sentiment analysis means understanding how a company’s data, products, and content are perceived by HCPs over time. Understanding this sentiment over time can aid in measuring how the needle is moving after implementing different scientific initiatives. Examples could include: a) how to engage HCPs depending on their sentiment, and b) which types of content are resonating with HCPs. Having a computer do this leads to consistent results.

ai and medical insights reading content.jpg

Combining and Comparing Medical Affair Data Sources

Medical Information call records can be categorized using NLP and ML models to identify the type of inbound inquiries. This can then be compared to Field Medical questions and activities. This analysis could show that your field insights are efficacy-focused, but most HCP calls are asking about safety data. With this information automatically generated by NLP and ML, MSL leads can quickly adjust the Medical Strategy by preparing safety documentation for Field Medical teams to share with HCPs. 

Downsides of using AI for medical insights

Large training sets are required

An accurate and robust training dataset is the backbone of supervised ML. A large amount of data with the desired answer or recommendation is required. For example, if you want to train an ML algorithm to categorize an insight, you would need a minimum of 100 insights. This is on the very low end; having 1000+ would lead to more accurate results. To date, this has been the biggest barrier to implementing ML within Medical Affairs. 

Will never be 100% accurate

AI has come a long way, but there is still a lot of work required for computers to make recommendations like humans. This means there will be errors. Think of it as a supplementary tool to help you do your job. Not the end-all answer. 

Treating human disease is complicated

Molecules, pathways, biology, treatment landscapes, competitors, clinical trials, products, etc need to be taken into account when treating patients. And let’s remember that MSLs are highly trained, with years and years of education. They understand and pull together multiple complex pieces of information to make recommendations that improve the lives of patients. Getting a computer to do the same thing is an enormous task. Because of this Medical Affairs will always require a human to make a recommendation. However, we can leverage AI to get to the recommendation more effectively, feel confident about it, and save time.

Summary of AI, ML, and NLP


AI will be a game-changer for analyzing Medical insights. It is already saving MSL teams significant time by tagging and categorizing insights automatically. Gone are the days of an MSL lead laboriously tagging insights one-by-one. 

Automation of labor-intensive tasks and time-savings are the big benefits of ML. AI cannot determine what trends from insights mean and which actions to take next. This is where Medical Affairs teams and their experience come in. MSL teams can take the next step by analyzing the results of various ML models and deciding what the trends mean for an organization’s customers, products, and patients.

So to answer your questions if robots will take all the MSL jobs and take over Medical Affairs. The answer is absolutely not. But they can help by taking care of some of the boring and labor-intensive work. Plus they are a lot faster at it :).

machine learning medical insights c.png