The Five Steps Toward Machine Learning

Everyone in the business world wants to get artificial intelligence (AI) to work for them. The idea that a computer program can basically predict the future is enticing and exciting. What if you didn’t have to test a material to know if it would improve a medical device? What if you knew before implementing a change in the manufacturing line how much time it would save? What if you knew a change in a product component would double sales?

Those questions are all answerable. When talking about AI and machine learning (ML), we’re no longer discussing the theoretical scenarios science fiction is made of. We’re talking about what’s possible, and happening. That’s an amazing thought, but even more amazing is how few life sciences companies are taking advantage of these technologies. Many want to use AI/ML applications, but few have the means to do so. The process is a long one, so the best approach is to take one step at a time.

1. Go Beyond the Basics
As the saying goes, “The first step is admitting you have a problem.” For most companies, the problem is their data is trapped. Whether that’s on physical paper or Excel spreadsheets is irrelevant. The point is, the data is in a place where it can’t be analyzed easily. They haven’t truly digitized their systems. In this stage of the process, you might be able to get some information out of the data. But it’s extremely limited and hard to get. You can also never be sure how dependable it is because, without a centralized location housing the data, you can never be completely sure you have all of it.

A good example is complaint data. Analyzing complaint data might seem straightforward—it quickly becomes complicated when you consider how many inputs there are. For simplicity’s sake, let’s say you have 10 facilities around the country and each has five employees that record complaints. You also have a website where consumers can register complaints and sometimes complaints are emailed to random employees. Even if you consolidate that data and eliminate duplicate complaints, you still won’t know which device has the most complaints or what can be done about it.

2. Descriptive Analytics
Fortunately, the next step in this progression does indicate which product has the most complaints, but not what can be done about it. In the most basic sense, you can manually compile the data and plug it into a spreadsheet to determine which device has the most complaints associated with it. However, this is much easier if you’ve invested in a good quality management system (QMS) that can consolidate complaints. With everything in a centralized location, the data is much easier to access and analyze. Even then, it will probably be limited in its abilities.

Running with the same example, descriptive analytics can indicate which device has the most complaints or what the most common complaint is. But that’s where its usefulness ends. This type of analytics leaves you guessing as to why that device has so many complaints. And it certainly can’t give suggestions as to how to reduce those complaints. Descriptive analytics tell what happened, but not why or how to proceed. Since your products are literally saving lives, these are important questions to answer.

3. Diagnostic Analytics
Taking things a technological step further tells us the “why?” of the problem. As the name implies, diagnostic analytics get to the heart of why the problem is happening. This is when the real benefits of analytics start to come into play and AI makes its debut. Besides explaining why something is happening, AI can show where the problems are, so you know where to focus your efforts. Natural language processing (NPL) can analyze complaint data to identify themes that wouldn’t be apparent using the analytics available in steps one and two.

The traditional complaint management system has a human enter information into a form. This inherently restricts the information that can be gleaned from the complaint. Human error and limitations in the form lead to misdirection when companies try to determine areas in which they can improve. NPL looks at the entire complaint to analyze what went wrong and why. It can also interact with a larger digital system to give a precise, root cause of the complaint. Diagnostic analytics serves as a transitional step from retroactive quality to proactive quality. It does look at the past, but makes suggestions on how to improve in the future. Of course, this is just the tip of the iceberg when it comes to analytics.

4. Predictive Analytics
Once you feel comfortable in diagnostic analytics, it’s time to quit living in the past and look to the future. Predictive analytics provide invaluable insights into a business. Forecasting and simulations let you see how decisions are likely to play out without implementing them. This level of analytics requires a level of digitization and interconnectivity that few life sciences companies have achieved at this point. Predictive analytics will only be as accurate as the data that’s being analyzed. If you’ve already adopted software as a service solutions, you’ve taken the first step. Getting those solutions to integrate or switching to ones housed on the same platform will give the baseline needed for predictive analytics.

Let’s say you’ve used diagnostic analytics to figure out that most of the complaints relate to devices that include a specific component. Changing that component is more complicated than a simple swap, however. The manufacturing line has to be altered and there are multiple ways this can be done. How do you know which option will be the most efficient and productive? Using predictive analytics, you can determine the answer to those questions before making the changes. That way you won’t have to make any costly changes only to determine later that even more changes need to be made. 

5. Prescriptive and Integrated ML
Predictive analytics are based on actions suggested by a user and represent an approach more efficient than trial-and-error. You can simulate how an action will affect processes, but predictive analytics don’t tell you what actions to take. Since solutions can be elusive or you might settle on a less-than-optimal solution, having a form of AI that suggests a solution is invaluable. Even better, the AI can automatically implement the solution without needing human interaction. ML can look at data and calculate risk more accurately and faster than any human, making ML-based decisions the most likely to succeed in a business.

Using predictive analytics to make the component change to a manufacturing line would be easier than trial and error, but still difficult. You’d have to code every scenario and run it through a simulation to arrive at the most efficient decision. ML does all this automatically. It learns as it goes and incorporates that information into future decisions. The more it analyzes, the more effective it becomes and the greater asset it is to your business.

The potential of AI to transform a medical device company is exciting, so you might be tempted skip a few steps to get to the most advanced step with the highest payout. The problem with this is without a foundational understanding of AI and the tools already in place, there is no way to effectively jump on board the ML train at the last stop. Skipping steps in a process as complicated as this one is a recipe for disaster. You could immediately invest in AI/ML applications, but if you haven’t taken the time to build the proper framework, the applications won’t produce effective results.

It’s worth it to slowly build up your AI/ML initiatives in your company. This can take years, but it’s better to do it right than to rush through it. Where you start depends on where the company is. Do you already have digital initiatives in place that will help take advantage of these technologies? Or are you still working on getting rid of paper? Regardless, take it one step at a time and you’ll find yourself with a competitive advantage that will put you at the top of your industry. 

---

Sue Marchant is the director of product, machine learning and artificial intelligence at MasterControl, where she is responsible for infusing the advanced data mining and analytics capabilities afforded by these technologies into MasterControl’s cloud-based product life cycle excellence software for the life sciences. She spearheads initiatives around actionable, predictive insights and optimized efficiency, productivity, and compliance.