top of page

Predictive Analytics: What are the Hurdles?

Updated: Aug 18

Predictive analytics has emerged as a powerful tool to forecast outcomes and make informed decisions in various fields. At EPCON, we are particularly interested in applying predictive modelling to infectious disease control and various other challenges in the health field. While these models offer excellent opportunities, they also come with a set of challenges. This piece looks at how to navigate through these obstacles to harness the true potential of predictive analytics.

Incompleteness: The Foundation of Accuracy

Predictive models rely on accurate datasets. Deficient models are often the product of incomplete or inaccurate datasets. Small datasets may not have good generalisability. To train neural networks, you need large datasets - but these can be very expensive.

EPCON uses Bayesian networks. These simpler, directed graphical representations can actually reason effectively in uncertain environments, which means they can find generalizable patterns. Bayesian networks are also robust, as they have greater capacity to see the “big picture” rather than getting stuck on details that may not be relevant. In essence, Bayesian networks are ideal in uncertain environments, like disease epidemics.

Predictive Analytics

Data Myopia: Expanding Horizons for Better Insights

Another challenge is related to the problem of inaccurate datasets. What happens if data is not representative: for example, the subjects used in the model might be classified in ways that are too limited? Under these circumstances, the conclusions drawn could be too narrow. One potential solution is to use the model with unseen data in order to identify and correct any hidden biases before they become a problem.

Narrowisation: Unravelling Complex Behavioural Patterns

Predictive models might impose artificial boundaries that restrict the range of anticipated behaviours in clients. This is especially relevant when it comes to health-seeking behaviours which can be complex and wide ranging. One way around this challenge is to use domain experts to identify relevant data sources and outcomes of interest.

Spookiness: Striking a Balance Between Tracking and Privacy

Clients, legitimately, become concerned about being tracked and part of a large automated system. This concern becomes even more prominent when dealing with stigmatised diseases like HIV and TB, necessitating strict assurance of anonymization and privacy protection during model training.

Skills: The Multidisciplinary Nature of AI

AI is a rapidly-moving field requiring a wide range of technical skills and expertise - for example, statisticians for model accuracy, data scientists, data engineers who understand model selection and evaluation, and others. It is frequently challenging to find the required skills. Collaboration across disciplines and continuous education is essential for team members to stay updated with the latest developments and ensure successful AI implementations.

Adoption: Bridging the Gap Between Tools and Users

Despite user-friendly platforms, there is resistance to adopting predictive analytics models. Distrust of metrics and attachment to traditional tools like Excel contribute to low adoption rates. Explaining "black box" models like neural networks poses an additional challenge, making it necessary to develop visually interpretable models like Bayesian networks that clarify how predictions are made. EPCON’s use of visually appealing and simple-to-understand dashboards helps a lot in this regard.

By addressing these hurdles, predictive analytics models have the potential to revolutionise decision-making and forecasting in various domains. In the area of health, predictive models can help tremendously to make informed decisions, control infectious diseases and strengthen healthcare systems.

bottom of page