The Human-Machine Matchup

April 1, 2019
Combining human reasoning with the machine learning functionality found in advanced analytics applications improves pharma processes

When human intuition and experience engage with machine learning and other advanced technologies, therein lies great opportunity for innovation. Subject matter experts bring the power of intuition and thoughtful responses critical for constructing an effective data analytics strategy. This is particularly true because human experience extends well beyond the capabilities of machine learning.

However, since analysis can suffer from human biases and heuristics, it is imperative to also take advantage of machine learning algorithms to identify patterns worthy of further exploration. Since having the means to describe one’s own bias is not sufficient to overcome it completely, the company culture should foster collaboration, particularly across varying viewpoints and experiences within pharmaceutical manufacturing.

In “Thinking, Fast and Slow,” the notable psychologist, Daniel Kahneman, wrote: “Intuition is nothing more and nothing less than recognition.” This helps explain why the most effective solutions combine human and machine learning capabilities to get the best of both worlds.1

This article provides a path forward to realize improvements, beginning with the acknowledgement of the challenges and culture changes required. Winding through the mindset required for optimal technology selection, we ultimately can create a vision of how a joint human- and machine learning-supported solution may be implemented successfully.

Background: The Challenge

On the culture front, many pharma firms are joining together in consortia-type models to leverage the tremendous opportunities for connecting human expertise with machine learning.2,3 Of particular interest is understanding how to use machine learning to extract valuable information out of the vast amount of raw chemical data in the drug discovery process.

On the technology front, the ability to utilize machine learning algorithms is becoming a reality. In the pharmaceutical and other process industries, interest in these approaches has been fueled by recent advancements and improved accessibility to machine learning functionality.

Machine learning has been described as “the science of getting computers to learn and act like humans do, and improve their learning over time in autonomous fashion, by feeding them data and information in the form of observations and real-world interactions.”4 Commonalities between the human mind and machine learning technology include:

  • Drawing on compiled patterns
  • Executing responses and generating intuition
  • Taking the opportunity to learn regularities through practice
  • Needing to experience variability while learning
  • Approximating a large number of items
  • Understanding the reasoning behind the intuition
  • Seeing patterns in randomness

To meet the stringent regulations placed on drug development and production, companies need to be able to work with transparent algorithms so their subject matter experts (SMEs) can understand and explain the causal reasoning behind the conclusions. Machine learning algorithms do not always arrive at the answers for the same reasons people do, which typically means one can’t inherently understand the outputs of machine learning systems, or directly trace their relationship back to the inputs. These algorithms rely on recognition, not reasoning, so it’s unlikely a random output from a complicated “black box” technique can be explained to any decent level of satisfaction, or relied upon without further inspection.

Therefore, it is necessary to assemble several easy-to-use tools for placing machine learning techniques directly into the hands of SMEs. Data science expertise can be leveraged through these tools to expand these resources broadly across the organization.

However, while progress is being made on the bulleted list above, there is still a broader challenge to realize the full potential for pharmaceutical manufacturing found at the intersection of technology, machine learning and human experience. These challenges include:

  • Making better decisions regarding data governance
  • Structuring datasets
  • Providing transparency of algorithms so conclusions can be understood and explained
  • Recruiting data science talent
  • Providing access to data silos, both within teams and across company sites
  • Streamlining data acquisition and process analytical technology (PAT) measurement initiatives
  • Making cultural changes to adopt new approaches

The emergence of pilot programs and scaled-up applications clearly shows some progress by companies, from assessing and reviewing the potential of innovative technologies to moving towards actual implementation. An example of this is shown in Figure 1, which illustrates this approach applied to creating more economical facility operations by improved heat exchanger monitoring.

Strategy — Coupling Human and Machine Power

While machine learning is an important element, it is not an end or a solution in and of itself. Instead, it is a tool SMEs can use to derive deeper understanding of pharmaceutical manufacturing processes. Figure 2 shows just how this can take place by empowering SMEs.

Machine learning models can make predictions and recommendations by finding patterns in data at a scale humans are incapable of handling, but there is no transparency, an SME is often not able to explain the how or the why. Therefore, machine learning must be combined with other important elements including:

Data Sources and Data Management: Pharmaceutical manufacturing companies are deploying smart devices to gather data not previously available. Process measurements, observations and data source connectivity are all required to obtain the breadth of data needed to build successful models. As a result, companies are now faced with the growing challenge for providing connectivity of these data sets with other manufacturing process data.5 Pharmaceutical companies are often saddled with old legacy systems, often from multiple acquisitions, containing disparate data. Many strategies rely on finding scalable solutions capable of connecting to these various systems.

Data Science and Mathematics: Statistics and other modeling approaches often follow widely accepted and well-documented best practices for building models and identifying important variables. While these simple, linear statistical models often provide an illusion of interpretability, they are often misleading. Using complex but more accurate algorithms is often a better approach. The need for SMEs in this process is clear because they must develop these complex algorithms, which can then be implemented to provide broader use.

Culture and Process: Strong leadership and cross-functional team structure both play important roles in facilitating organizational use of the data, models and complex algorithms developed by SMEs. Many automated manufacturing sectors are replacing or enhancing statistical methods with machine-learning techniques to improve efficiency, reduce cost, optimize product yield and enhance product quality. Within pharma, detailed standard operating procedures exist to help ensure processes are robust, stable and yield predicted quality. These procedures provide a structure for using machine learning and other techniques.

Human Intuition and Reasoning: With respect to the important element of leveraging human intuition and reasoning, Kahneman published a masterful theory of human decision-making and judgement, which operates under two inter-locking systems. In this theory, “System 1” is our rapid and largely unconscious mode of operation, drawing largely on association and metaphor — a “gut reaction” way of thinking. “System 2” is our unrushed, analytical, deliberate and conscious mode of reasoning — a “critical” way of thinking.

System 1 operates well in many cases, but there are times when it can introduce errors of bias since it seeks to create a coherent and plausible story by relying on assumptions, memories and pattern-matching. When the thought process uses only existing evidence and ignores absent evidence, System 1 can create a believable story, but often not a complete solution. System 1 limitations often lead to cognitive biases, or unconscious errors of reasoning. A few examples of where these errors in judgement often show up include the following:

  • Laws of small numbers
  • Assigning cause to random chance
  • Illusion of understanding
  • Hindsight bias

But there is a better path, one which involves the application of System 1 and 2 together.


Because machine learning algorithms rely only on recognition without reasoning, and because the human brain has its own challenges regarding bias, the best solution is to:

  • Understand the power and limitations of machine learning algorithms
  • Optimize a system to promote effective use of human intuition and reasoning
  • Provide the system with a user-friendly interface to take advantage of both machine learning and human reasoning

Further, the approach of leveraging teamwork within a software environment helps proactively identify and eliminate many human biases. In conjunction with machine learning algorithms, the impact from this broader human engagement will then result in a better outcome.

It is important to identify and implement an advanced analytics solution that effectively couples human and machine power to:

  • Structure datasets to maximize impact by delivering better error tracking, guiding predictive maintenance activities, and improving supply chain monitoring to reduce material costs and increase product yields.
  • Leverage embedded machine learning algorithms to draw out profiles and trends which can inform and guide.
  • Create and share step-by-step analytics processes and assemble them into repeatable documented workflows. Validated analytic workflow processes can be assembled into repeatable documented workflows, which can be utilized by the broader team to extend use and achieve better quality with lower risk.
  • Draw in multiple people to reduce the impact of human bias and increase the opportunity for more thoughtful and deliberate data analysis.

For machine learning algorithms to be effective within this solution, it is important to understand the features used to describe the domain-specific data, obtain adequate data to train the models in the first place, and set aside a portion of the training data set for cross validation.

Business rationales for using coupled human and machine learning include (Figure 3):

  • Diagnostic: Avoiding downtime and affiliated losses through root cause analysis
  • Predictive: Increasing revenue by maintaining nearly 100 percent uptime at lower cost
  • Monitoring: Delivering advisory real-time and prediction view of process and asset status
  • Prescriptive: Evaluating options to make decisions that optimize outcomes
  • Descriptive: Achieving near real-time distribution of insights to share and inform plant-wide decisions

The Value is Clear

Artificial intelligence and machine learning technologies promise to revolutionize traditional pharmaceutical practices, so companies need to align cultural best practices to harness the best of both humans and machines. Breakthroughs will occur when pharmaceutical firms effectively leverage cross-company human intuition and experience to reduce bias and increase the likelihood of a better outcome. To innovate fully, SMEs need:

  • The ability to interact with the data and correlations to further apply their rich knowledge and intuitions based on experiences not captured within the database.
  • Quick access to a data-rich environment where machine learning algorithms can be applied to delve into the data in a non-biased manner to draw out currently unrecognized correlations.
  • Opportunities to explore the processes with colleagues to further eliminate biases and heuristics that can often cloud or even mislead the user.

There are compelling reasons why proven successful machine learning techniques should be used to improve validated manufacturing processes. Ample experience and consensus around best practices exist, and the value is clear: faster time-to-market; more scalable, robust and reliable processes; and higher quality products — all at a lower cost.  


D. Kahnman, “Thinking Fast and Slow,” Farrar, Straus and Giroux, New York, 2011.
K. Sennaar, “AI in Pharma and Biomedicine – Analysis of the Top 5 Global Drug Companies,” Emerj, Dec. 12, 2018.
S. Koperniak, Department of Chemical Engineering, ”Applying Machine Learning to Challenges in the Pharmaceutical Industry”, MIT News, May 17, 2018.
D. Faggela, ”What is Machine Learning?”, Emerj, Dec. 21, 2018.
J. Schmidt, “How Big Data Is Transforming Pharmaceutical Manufacturing”, Pharmaceutical Online, Aug. 12, 2016.

About the Author

Lisa Graham | PhD and PE