AI Brings New Opportunities to RWE Research – Its Use in Health Data Analysis is Limited

Artificial intelligence (AI) is transforming research and working life across various industries. But what is its significance in health data analysis and real-world evidence (RWE) research? Medaffcon’s experts, Juhani Aakko and Lisse-Lotte Hermansson, share their perspectives on AI in their work.

Medaffcon provides research, expert services, and consultancy for the pharmaceutical industry and healthcare sector. One of its core areas of expertise is real-world evidence (RWE) research based on healthcare data. 

Juhani Aakko, senior data scientist, works on RWE research in Finland. In his role, Aakko utilizes generative AI for tasks such as information retrieval and presentation preparation. However, the use of generative AI in RWE data analysis is limited due to the strict requirement that the data must be processed in secure environments with limited internet connection. 

“We analyze healthcare data in Findata’s secure environment or HUS Acamedic.  There is very limited access to the internet and the data cannot be removed from these environments to ensure the protection of health information and prevent the risk of patient data breaches,” Aakko explains. 

Aakko believes generative AI could play a significant role in healthcare, particularly in automating time-consuming documentation tasks. Automating structured data entry could ease professionals’ workloads and improve data quality. 

Machine Learning and AI in Medaffcon’s Project 

Artificial intelligence (AI) is a broad concept that is not limited to generative AI. Medaffcon has conducted a groundbreaking registry study using machine learning and AI. The study analyzed all surgeries performed in the Helsinki University Hospital (HUS) area between 2015 and 2019. The findings showed that current and former smokers face a significantly higher risk of complications after surgery. 

The dataset included about one million surgeries. Initially, clinicians reviewed 20,000 patient records and classified patients as smokers, quitters, or non-smokers. This classification was then used to train a machine-learning algorithm. 

By applying machine learning and AI, smoking data for over a million surgeries was identified. The algorithm classified smokers, quitters, and non-smokers from the patient records. Physicians predefined the complications to be analyzed; ultimately, 158,638 surgeries were included in the AI analysis. 

 “Machine learning and algorithms enabled a project that would have been practically impossible to execute in the past,” Aakko says. 

According to Medaffcon’s Country director for Sweden, Lisse-Lotte Hermansson, AI offers significant advantages in text-based data analysis when it is feasible to use. 

 “Traditional data cleaning can take up to 80 percent of the total analysis time, but with AI, this process becomes much easier,” Hermansson notes. 

Machine Learning Models for Classification and Prediction 

Machine learning models are particularly valuable for classifying patient groups and making predictions. 

“We haven’t used them extensively yet, as clients usually have clear definitions of the patient groups they want to study,” Aakko explains. 

However, analyzing sufficiently large datasets could help to find patient groups who might benefit more from a specific therapy. 

“This is an interesting possibility, but it’s unlikely to be very beneficial. Much is already known about drug mechanisms, so the chances of finding groundbreaking insights through data analysis are unlikely.” 

AI Must Operate with the Highest Ethics and No Bias 

Aakko and Hermansson have few concerns about AI or its regulation in RWE research.  

“The possibilities for processing data are currently strictly limited, so I’m not worried at the moment,” Aakko states. 

Hermansson emphasizes the importance of properly developed prediction models. 

“Naturally, AI algorithms need to operate with the highest ethical standards and no biases. They must be trained with relevant patient populations, using excellent data quality, and applied with critical thinking. The interpretability of AI models is crucial for understanding the insights and predictions they generate, which can be challenging given the complexity of deep learning models,” Hermansson explains. 

She further notes that identifying individuals from large health datasets is almost impossible. 

“This should not be a primary argument against utilizing untapped information to improve patient outcomes.”