29.07.2020 / Artificial intelligence in clinical research

 

Artificial intelligence (AI), machine learning (ML), artificial neural networks (ANN) or deep learning (DL) are just some of the buzzwords one seems to catch almost everywhere for a while now. This totally makes sense since highly efficient and precise automatization and / or decision making is appealing for lots and lots of fields and applications, hence also for clinical research.

As big as the still increasing interest in AI, as diverse are the applications an artificial intelligence can be developed and used for. Therefore it is not overwhelmingly surprising, that, through the process of developing an AI for a specific use, you might come across further possibilities to optimize currently implemented processes.

This is a first glimpse into our AI developments at Alcedis and of how we optimize our systems and expand our capabilities along the way. Let’s start with a few basics…

 

What is AI anyway?

 

This already is a question not easy to answer. Loosely speaking we are talking about AI whenever a device or algorithm perceives its environment to make decisions and take actions that maximize its chance of successfully achieving goals such as recognition, planning, and / or learning, just to name a few.

The issue with the definition of artificial intelligence is that, as soon as a new problem has been solved and a new goal of AI research has been achieved, the respective program or algorithm is not considered as AI anymore. One example of this is the traffic sign recognition software used in modern cars. Even though it is a complex computational task “just checking if one picture looks like another does not require real intelligence”. This is the so-called AI effect and leads to another widespread definition of artificial intelligence: AI is whatever has not been done yet.

Of course, these two definitions do not cover the entire field of AI research, and for some parts there are precise definitions. However, they give a good impression of the ubiquitous struggle of whether it is correct to designate a new program or algorithm as AI or not.

Our goal for AI at Alcedis will actually be a self-expanding software for complex pattern recognition and evaluation.

 

What does AI have to do with clinical research?

 

Well, a short answer would be “A lot! Especially with statistics.”, but let’s go for a little more detail. Clinical research is based on data documenting treatment procedures and courses for individuals, and what renders this data and its evaluation quite complex is the word “individual”.

Valuable results require clean data and thus good data management, which is a time-consuming and complex task that continuously accompanies the process of data documentation of a study. However, the analysis of survival data, the search for correlations between different characteristics of a subject, his or her therapeutic success or failure, concomitant medications, adverse events, etc., all this takes place at the very end, according to a beforehand defined analysis plan. There may be some statistical reports or even interim analyses along the way but most commonly those do not cover every aspect of the final analysis, often because this is simply not feasible. For this reason things might get quite busy close to the end of data documentation, when data management and statistics get the closest together. Good specifications and data review schedules help enormously during this time, yet it remains a source of errors. Now this is where a good AI can do its magic by not only assisting during this time but by mainly preventing it from even happening.

A data management system that is given this tool becomes even more agile and reliable, which results in optimal study conditions due to maximum reliable and clean data.

 

First steps in developing an AI

 

With all those benefits in sight what are we waiting for? Well, unfortunately it is not that easy. Developing an AI does require lots of preparatory work. And even with a detailed specification of its desired capabilities there is no button labeled “Build AI” that you just push. Meaning there is not the one method of how to bring it all together. But let us save this for later and focus on some of the preparations.

How detailed is that kind of information usually available? How reliable is it? How reliable and complete can data be anyway when it cannot be measured but comes from a subjects’ memory? Do we need to weight information differently and if yes how? What does this information correlate with and can we extract the same kind of information value elsewhere? The list of questions goes on and it does so for every single part of the puzzle.

By analyzing the data and finding answers to as many of those questions as we can, we are steadily improving our final product already and identify the most important set screws for a later optimization.

 

First benefits along the way

 

Even through the process of data preparation and analysis is one of the first steps in the development of an AI based on statistical methods, benefits and knowledge you may have missed out on so far can emerge and these should not be underestimated. Speaking from experience we can provide an example as easy as useful.

Often, when an individual shall document a categorical variable, for example some kind of medication, in an eCRF this is done with a dropdown menu from which the desired answer is to pick. If none of the preset options fits, there is usually one more option, called “other” or likewise. This “other”-option triggers some additional field were free text can be entered.

For a medication it might happen that it is overlooked in the dropdown and then documented as a free text under “other” instead, maybe even with a typo. We can apply an automated search algorithm to the respective free texts, comparing the hits to the original dropdown, or any other desired library, and then forward a message to a data manager for confirmation. Additional implementation of a so-called fuzzy match does even detect medications with typos up to a certain degree.

As promised this example is pretty straightforward and only logical, yet there are a lot of data documentation systems used in clinical studies were such additional checks are not performed and some data may be lost.

If this topic caught your attention please don’t hesitate to contact us and make sure not to miss the next update of this series that will take us on a journey with some highlights in the current development and implementation of AI into clinical studies at Alcedis.