Clinical trials represent the most important part of drug development costs. In some cases, drug research and development costs can reach $2 billion per drug, with long timelines of 10 to 15 years. Pharmaceutical companies are turning to AI to shorten discovery timelines while significantly maintaining rigorous standards.
AI-driven approaches have demonstrated significant improvements in efficiency and diagnostic performance in controlled clinical research environments. However, there is growing evidence that success in trials does not guarantee real-world impact.
A 2025 peer-reviewed narrative review highlights a persistent gap between AI's superior performance in tightly controlled environments and its variable effectiveness in routine clinical practice, where adoption is often limited by scalability, workflow integration, data heterogeneity, and governance. Therefore, to realize the full potential of AI, it is important not only to have technical validation but also to responsibly integrate these tools into everyday healthcare systems.
In a controlled oncology setting, an AI-driven mortality prediction system increased critical illness conversation (SIC) rates from 3.4% to 13.5%. However, the same body of evidence indicates that such benefits often do not generalize to heterogeneous real-world clinical environments due to workflow, data, and scalability challenges.
Emerj Editorial Director Matthew DeMello sat down with Bayer's Patricio La Rosa on the AI in Business podcast to continue their earlier conversation and discuss the role of AI across the clinical spectrum, from trial design to long-term efficacy tracking.
The following article will highlight three key takeaways from their conversation.
- Establishment of scaling sensing modality: Build AI-driven diagnostics that enable a seamless transition from R&D to clinical practice without incurring prohibitive computing costs.
- Avoid regulatory and adoption bottlenecks: Leverage probabilistic AI to identify robust biomarkers faster while recognizing that approval timelines and patient engagement hurdles remain the same.
- Cybersecurity and data monetization: Explore new incentive models to protect patient privacy amidst growing threats and help patients share in the value their data creates.
Listen to the full episode below.
guest: Patricio La Rosa, Head of End-to-End Decision Science in Seed Production Innovation, Bayer Crop Science
Expertise: MLOps for industrial analysis, biostatistics, and biophysical modeling
Easy 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, we are leading a global decision science initiative that incorporates AI into seed production, supply chain planning, and manufacturing to create significant business impact. His work spans industry and academia, includes peer-reviewed research, extensive teaching at Washington University in St. Louis, and focuses on building scalable and responsible AI systems.
Establishment of scaling sensing modality
Host Matthew DeMello begins the conversation with questions about where and how AI is being applied in clinical trials, not just for basic efficiency and automation, but to capture meaningful data and insights to improve future clinical trials. La Rosa responds by emphasizing that, regardless of the sensing method used, there is always an upfront base cost for infrastructure.
He details how the initial costs of drug development can be focused on estimating a drug's specific effectiveness. However, once a drug is developed and its effectiveness confirmed, the focus changes. He went on to explain that the new challenge is determining how to reliably continue to measure and confirm the effectiveness of drugs in real-world use.
Even as AI allows us to analyze more data than ever before, we need to enable scalable technology that supports good clinical practice and economic benefits within reasonable economic constraints, according to La Rosa. He warns Emerj Executive Podcast listeners about the need to pay for computationally intensive clinical models during research and development.
La Rosa concluded by emphasizing the importance of sensing modalities not only being scientifically valid in research settings, but also industrialized for use in medical settings without imposing unsustainable costs on patients, healthcare providers, and even insurance companies. He gives an example of the use of fMRI, which may work well during development but becomes too expensive for continuous patient monitoring in practice.
Overcoming regulatory and adoption bottlenecks
La Rosa went on to explain that even with dramatic advances in biomarker discovery, the regulatory pathway remains the same. Regulatory reviews take time, and they happen for a reason. Even as regulatory pathways speed up as drug discovery capabilities improve, companies are still limited in their ability to find and recruit patients.
According to La Rosa, these problems don't even need to be resolved yet. He believes it is important to prioritize leveraging technology first before addressing how regulatory processes evolve as technology advances and matures.
La Rosa agrees that, despite the differences, many different technologies and approaches are often referred to as AI, and that differentiation can help increase transparency and gain buy-in from the business side. Overall, he considers deep learning techniques as AI and groups the rest under machine learning and statistics.
“We tend to use the term AI for different things, even though they’re not exactly the same thing. I think it’s important to differentiate because that transparency leads to trust and buy-in from businesses.”
Today, I personally reserve “AI” as a deep learning technology and group everything else under machine learning and statistics. At the end of the day, that clarity doesn't matter all that much to patients, but it matters a lot internally as organizations decide how to responsibly deploy, manage, and scale these systems. ”
– Patricio La Rosa, Head of End-to-End Decision Science in Seed Production Innovation, Bayer Crop Science
La Rosa further explains that most patients won't mind the full use of AI to refer to seemingly disparate technologies. Ultimately, he concludes, patients want a solution to their problem and are interested in what will work, especially if they are desperate for a solution.
Cybersecurity and data monetization
When asked how security considerations relate to patient targeting in clinical research workflows and how this impacts the broader system, La Rosa emphasized the responsibility that all R&D workflows and critical assessments have to ensure patient data is protected.
La Rosa also explains that the better they become at analyzing information, the more adept attackers will be at circumventing barriers. According to La Rosa, the arms race he describes has made it imperative that advances also be made in the field of cybersecurity.
The conversation turned to the possibility that patients could eventually expect financial compensation for their data and thereby contribute to advances in treatment. LaRosa positions this as an investment in treatment, suggesting that patients can participate and benefit from the value created by the data. He also highlights broader ethical issues about how to balance individual ownership of information with collective contributions to health and human progress.
