Practical Application of a Single Layer Perceptron and the Path to Unlocking the MSL Value Algorithm (MVA)

BACKGROUND

Artificial intelligence (AI) is changing how the medical affairs industry evaluates and predicts traditional medical science liaison (MSL) value-driven metrics. This paradigm shift is occurring as a result of the increased learning capabilities of neural networks. The implementation of artificial neural networks (ANNs) in medical affairs workstreams will enhance MSL performance by providing dynamic insight into value-driven field activities that complement the broader organizations business model. This technological advancement necessitates the need for a better understanding of how these neural networks operate.

Understanding the basic architecture and mechanism of how a neural network functions can help us in adopting a new data driven approach that can merge quantitative and qualitative MSL metrics. This approach will enable medical affairs teams to optimize their field activity to bring the most value and impact from a scientific and business perspective. This can be accomplished by developing what we refer to as the MSL Value Algorithm (MVA).

The customized MVA will be capable of evaluating and maximizing every unique Medical Affairs key performance indicator (KPI) with the help of deep learning technology. The data-driven approach of using an MVA could help position MSLs more as a complementary and cross-functional business partner within their organization by showcasing their unique business acumen and domain expertise. Our journey begins with an understanding of the basic unit of a neural network, the perceptron. We then show how the practical application of a single layer perceptron can be used to build a more complex deep learning algorithm called the MVA.

NEURON V. PERCEPTRON

A neural network is analogous to our human brain. Fundamentally, a neural network consists of a neuron. Several neurons arranged in an interconnected structure makes up a neural network. Each neuron is linked to the inputs and outputs of nearby neurons (refer to Figure 1). In the context of AI, a perceptron is analogous to a human neuron.

Figure 1. A neural network in AI models the neural network in humans

The perceptron was devised in 1957 by Frank Rosenblatt at the Cornell Aeronautical Laboratory. It was meant to mimic a human neuron in its structure and function. A perceptron functions as follows:

SINGLE LAYER PERCEPTRON & MSL METRICS

A SLP is depicted in Figure 2 below. The activation function is a special type of mathematical function that helps execute a result. In our example above we will use a random # of KOL interactions/month and # of medical insights/month. These random metrics are weighted by the organization (i.e., higher weighting is more important to the needs of the organization). After the additive sum of the weighted values are calculated, this corresponding value is then entered into an activation function. The value generated by the activation function is compared to a threshold value set by the MSL leader.

For simplicity, we will use a basic stepwise function to describe how a SLP can be used as a neural network to compute the value and impact an MSL had for a month in the field. It is important to note, additional input values (Xi) and weights (Wi) can be entered, but we are limiting this example to two variables to show the practical application of a SLP or neural network.
In our example, an arbitrary predefined threshold activation value set at 18 would equate to a threshold value of ‘1’ being generated by the neural network. This would correlate to the MSL meeting their monthly value metric. Conversely, a value below 18 would indicate the MSL did not meet their monthly performance metric.

Configuring our SLP in a machine learning application allows us to see how this approach could be implemented across medical affairs teams. We utilized our Microsoft® Azure technology stack and our internal integrated development environment (IDE) to execute our machine learning algorithm. It is not important to understand the Python code in the example below. Rather, this example highlights how a basic SLP can provide additional value to managers when evaluating MSL performance metrics. It is important to point out that these numerical metrics were randomly generated to demonstrate how this SLP computes these two values. No client data or third party data was used for our example.

Figure 4. SLP implementation in a machine learning algorithm

The output of the SLP machine learning algorithms can be provided in real-time if configured to an organization’s dashboard or CRM platform. This methodology would enable managers the opportunity to react efficiently to emerging trends coming from MSL field activities. It would also afford the opportunity to provide constructive feedback to the MSL team that would in turn, help optimize performance and bring greater value to an organization.

MULTI-LAYER PERCEPTRON & THE MSL VALUE ALGORITHM

Whereas a SLP has only one hidden layer (refer to Figure 3), a multi-layer perceptron (MLP) has multiple layers. This is also referred to as a deep neural network (DNN). Similar to an SLP, hidden layers receive input data, processing it using a specific weighting, and applying an activation function to produce a specific output.
A SLP can only learn linearly separable patterns, thus limiting its practical use. For non-linear data, such as qualitative MSL activities, a SLP is not sufficient for data analysis. In these instances, a MLP or DNN is preferred because it can process complex and massive data and perform more intense mathematical computations.

Applying a real-time data-driven approach to all MSL field activities, both qualitative and quantitative, can now be accomplished more efficiently with the help of MLPs and Deep Neural Networks. For example, we can take a list of popular MSL field activities from a spreadsheet and enter them in a MLP to yield customized performance results for an entire MLS team in a matter of minutes (refer to Figure 6).

A customized MSL Value Algorithm can be created using the principles of a MLP and leveraging the power of current deep neural networks. This approach has the potential to optimize workflows, enhance productivity, and bring maximum value to MSL field contributions.
We will provide a more detailed explanation of configuring a deep neural network in an upcoming publication. Please note that configuring a robust DNN requires additional considerations that are not addressed in this publication. These include, but are not limited to, backpropagation, bias, cross entropy, other activation functions, and the vanishing gradient problem. These considerations have been omitted for the purposes of simplifying our basic model and to highlight the feasibility of using current MSL KPIs and computing power to assess value using the power of neural networks.

CONCLUSION

Deep neural networks can augment traditional MSL KPI assessments to help provide additional depth and dimension to the incredible work MSLs provide to an organization. Both quantitative and qualitative MSL contributions can now be evaluated using deep learning techniques. This approach allows organizations to see structure and correlations in data that would otherwise have been difficult to see.

Neural networks need two things to operate well: a high level of compute power and lots of data. We have both elements in abundance today, therefore making implementation of deep learning strategies for MSL activities feasible and impactful for MSL teams.

By partnering with internal data scientists, medical affairs leaders can explore new ways of evaluating traditional MSL KPIs using deep learning strategies. This data-driven approach can enable MSL teams to customize their field activities in real-time to help bring maximum value to their organization.

References:

About the authors:

Michael Shim, PharmD, Certified AI Scientist (CAISTM)

Michael Shim serves as a National Field Medical Director for the division of Deployment Solutions at Syneos Health. He is an innovative leader with specialized therapeutic experience in solid tumor oncology and is passionate about artificial intelligence and machine learning implementation in medical affairs.

Larry Dollar, PharmD, MS, MSL-BC

Larry Dollar is a Vice President of Medical Affairs at Syneos Health. He has also been a Senior Director and Field Medical Head at Takeda Oncology since 2014. Larry has also served in various MSL and MSL Director roles with Sanofi/Aventis, TAP, and Millennium Pharmaceuticals since 2000. He received his BS in Pharmacy from the University of Kansas, his MS in Pharmaceutical Sciences from the University of Missouri-Kanas City, and his PharmD from Idaho State University. His pharmaceutical career also includes experience in clinical research, medical research, and regulatory affairs. Larry currently lives in Prairie Village, Kansas, and enjoys spending time with his wife and four children.