Clinical trials are the most important part drug development costs. In some cases, pharmaceutical R&D costs can approach up to $2 billion per drug and come with extended timelines of 10 to 15 years. Pharmaceutical companies are turning to AI to shorten discovery times while meeting rigorous standards.
AI-based approaches have demonstrated significant gains in diagnostic efficiency and performance in controlled clinical research settings. However, evidence increasingly shows that successful trials do not guarantee real impact.
A peer-reviewed narrative in 2025 goodbye highlights a persistent gap between AI’s strong performance in tightly controlled environments and its variable effectiveness in routine clinical practice, where scalability, workflow integration, data heterogeneity, and governance often limit adoption. Realizing the full potential of AI therefore depends not only on technical validation, but also on the responsible translation of these tools into everyday healthcare systems.
In a controlled oncology setting, an AI-based mortality prediction system increased critical illness conversation (CIS) rates from 3.4% to 13.5%. Yet the same body of evidence shows that these gains often do not generalize to heterogeneous real-world clinical environments due to workflow, data, and scalability issues.
Emerj Editorial Director Matthew DeMello spoke with Bayer’s Patricio La Rosa on the “AI in Business” podcast to continue their previous conversation about the role of AI in the clinical continuum, from trial design to monitoring long-term effectiveness.
The following article will focus on three key points from their conversation:
- Establish scaling detection methods: Create AI-based diagnostics that enable a seamless transition from R&D to clinical practice without prohibitive computational costs.
- Overcoming regulatory and recruitment bottlenecks: Leverage probabilistic AI to identify robust biomarkers more quickly, while recognizing unchanged approval timelines and barriers to patient engagement.
- Cybersecurity and data monetization: Protect patient privacy in the face of growing threats and explore new incentive models where patients share the value created by their data.
Listen to the full episode below:
Guest: Patricio La Rosaresponsible for end-to-end decision science in seed production innovation, Bayer Crop Science
Skill: MLOps for industrial analysis, biostatistics and biophysical modeling
Brief recognition: Patricio is a leader in decision science and AI with over 20 years of experience applying machine learning and quantitative modeling to large-scale scientific and operational challenges. At Bayer Crop Science, he leads global decision science initiatives that have integrated AI into seed production, supply chain planning and manufacturing, resulting in significant business impact. His work spans industry and academia, with peer-reviewed research, extensive teaching at Washington University in St. Louis, and a focus on creating scalable and responsible AI systems.
Establish scaling detection modalities
Host Matthew DeMello opens the conversation by asking how and where they are applying AI in clinical trials, not only for basic efficiency and automation, but also to capture meaningful data and insights that improve future trials. La Rosa responds by pointing out that there is always an initial base cost for the infrastructure, regardless of the detection modalities used.
It details how the initial cost of drug development could be focused on estimating a particular drug effect. However, once the drug is developed and the effect has been detected, the focus shifts. He goes on to explain that the new challenge is how to continue to reliably measure and confirm the effect of a drug in real-world use.
According to La Rosa, while AI allows us to analyze more data than ever before, it must lead us to scalable technology that supports appropriate clinical practice and economic gain within reasonable economic constraints. He cautions Emerj’s Executive Podcast audience about the need to pay for clinical practice models that have a high computational load during R&D.
La Rosa concludes by emphasizing the importance of sensing modalities not only being scientifically valid in research environments, but also industrialized so that they can be used in healthcare without imposing unsustainable costs on patients, providers, or even insurers. He gives an example of using fMRI that might work well during development, but proves too expensive in practice for continuous patient monitoring.
Overcoming regulatory and recruitment bottlenecks
La Rosa goes on to explain that even with dramatic advances in biomarker discovery, the regulatory pathway remains unchanged. Regulatory reviews take time and are in place for a reason. Even as the regulatory process is accelerated to match increased drug discovery capacity, companies remain subject to limitations in their ability to find and recruit patients.
According to La Rosa, there is no need to even resolve these issues at this time. He believes it is essential to prioritize leveraging technology, before addressing how regulatory processes can evolve as it evolves and matures.
La Rosa agrees that many different techniques and approaches are often called AI, even though there are differences, and that differentiation helps build transparency and gain buy-in from businesses. Overall, it considers deep learning techniques as AI and groups the rest under machine learning and statistics:
“We tend to use the word AI for a lot of different things, even if they’re not exactly the same. I think it’s important to differentiate, because that transparency helps build trust and buy-in across the business.
Today, I would personally reserve “AI” for deep learning techniques, then group everything else under machine learning and statistics. Ultimately, this clarity matters less to patients, but it matters a lot internally when organizations decide how to responsibly deploy, govern, and evolve these systems.
– Patricio La Rosa, Head of End-to-End Decision Science in Seed Production Innovation at Bayer Crop Science
La Rosa further explains that most patients will not mind the widespread use of AI to refer to seemingly different techniques. He concludes that ultimately, patients are interested in what works because they want their problem solved, especially if they are desperate for a solution.
Cybersecurity and data monetization
When asked how security considerations intersect with patient targeting in clinical research workflows and how they affect the system as a whole, La Rosa emphasizes the responsibility that lies with every R&D workflow and critical assessment to ensure patient data is protected.
La Rosa also explains that the more information we analyze, the more skilled attackers will be at bypassing barriers. The arms race he describes also makes progress in cybersecurity essential, according to La Rosa.
The conversation has shifted to the possibility that patients could eventually expect financial compensation for their data, thereby contributing to therapeutic advances. LaRosa frames this as an investment in healing, suggesting that patients could participate and benefit from the value created by their data. It also highlights a broader ethical question: how to balance individual ownership of information with collective contributions to the progress of health and humanity.